Annual Report 2012 Himeji Initiative...Research Center, Osaka University, Osaka, and Himeji Initiative in Computational Medical and Health Technology, Hyogo, Japan. al. have successfully

Annual Report 2012

Himeji Initiative

in Computational Medical and Health Technology

１．研究者名簿

1

２．研究活動報告

（１）研究紹介と定例ミーティング

（２）センター内見学会

（３）他研究施設との交流

（４）受賞

1

３．第 2 回医療健康情報技術研究センターシンポジウム

（ICETET2012 Joint Symposium）

1

４．研究業績 1

医用生体情報解析グループ 1

生体マイクロセンシンググループ 1

看護アプリケーショングループ 1

生体計測シミュレーショングループ 1

ヘルスケアシステムグループ 1

生体数理モデリンググループ 1

５．競争的資金 1

６．Press 記事・メディア発表 1

医用生体情報解析グループ

畑豊教授（センター長）

藤井健作教授

佐藤邦弘教授

小橋昌司准教授

森本雅和助教

荒木望助教

生体マイクロセンシンググループ

前中一介教授

藤田孝之准教授

神田健介助教

才木常正主任研究員（兵庫県立工業技術センター）

看護アプリケーショングループ

新居学助教

湯本高行助教

坂下玲子教授（看護学部）

内布敦子教授（看護学部）

生体計測シュミレーショングループ

倉本圭准教授

林治尚准教授

下權谷祐児助教

齋藤歩助教

比嘉昌助教

佐藤哲也教授（シミュレーション学研究科）

内海裕一教授（高度産業科学研究所）

ヘルスケアシステムグループ

相河聡教授

畠山賢一教授

河合正准教授

山本真一郎助教

内田勇人教授（環境人間学部）

生体数理モデリンググループ

松井伸之教授（副センター長）

上浦尚武准教授

礒川悌次郎准教授

池野英利教授（環境人間学部）

2012.04.10 山本助教研究紹介『電磁環境対策のための電波吸収遮蔽技術』

2012.04.23 礒川准教授研究紹介

『これまでの研究紹介－生物に学んだ情報処理を中心として－』

2012.05.07 比嘉助教研究紹介

『これまでの研究紹介－Reverse 式人工肩関節の計測を中心として－』

2012.06.04 河合准教授研究紹介

『マイクロ波回路からマイクロ波エネルギー応用まで』

2012.06.18 佐藤邦弘教授研究紹介『ホログラフィによる生体医用光計測』

2012.07.02 林准教授研究紹介『水溶液や生体関連高分子の分子シミュレーション・

兵庫県立大学の情報システムの設計と運用』

2012.10.29 藤井教授研究紹介

『特性が未知のシステムを分析，解析する適応信号処理』

2012.11.12 荒木助教研究紹介

『これまでの研究紹介－BCI リハビリテーションシステムなど－』

2012.11.29 医用生体情報解析 Gr(畑教授) 今後の活動方針について

2012.12.12 生体マイクロセンシング Gr(前中教授) 今後の活動方針について

2013.01.07 生体数理モデリング Gr(松井教授) 今後の活動方針について

2013.02.01 看護アプリケーション Gr(新居助教) 今後の活動方針について

2013.03.05 ヘルスケアシステム Gr(相河教授) 将来計画について

2012.08.06 第１回センター内見学会が行われました

2012.08.29 第 2 回センター内見学会が行われました

2012.10.05 第 3 回センター内見学会が行われました

2012.06.15 講演会が行われました

Dr. Michael Rohan, ハーバード大学マクリーン病院

『fMRI の手法、データ、解析のトピックと fMRI の歴史』

Functional Magnetic Resonance Imaging: History and Methods

2012.07.09 共同研究提案会が行われました

大阪大学免疫学フロンティア研究センター森特任研究員

『高磁場 MRI を使った脳内免疫細胞動態追跡法(作用機序の解明に向けて)』

http://www.eng.u-hyogo.ac.jp/eecs/hata/hi-med/photo20120615.html

http://www.eng.u-hyogo.ac.jp/eecs/hata/hi-med/photo20120709.html

2012.11.07 ICETET2012 から Outstanding Contribution Award を頂きました

http://www.eng.u-hyogo.ac.jp/eecs/hata/hi-med/award.html

第 2 回医療健康情報技術研究センターシンポジウムを 2012 年 11 月に開催いたしました

ICETET2012 Joint Symposium （2012.11）

- 2012 Fifth International Conference on Emerging Trends in Engineering and Technology-

� Abstract—This paper describes a seminiferous tubules

evaluation system using an ultrasonic probe. This system evaluates diameters of seminiferous tubules for azoospermia patients. We employ a 5.0MHz ultrasonic single probe. In the experiment, we employ large and small nylon lines as the healthy and unhealthy seminiferous tubules. We acquired the waveforms by the ultrasonic probe and calculated frequency spectrum by short-time Fourier transform. The system visualizes the position of small and large line by a frequency map. The frequency map shows distance and frequency value. In the results, our proposed method detected the large and small lines. Additionally, the system evaluated distance between probe and these lines clearly. The proposed method thus successfully evaluated the position of the large lines.

Index Terms—Frequency, First Fourier Tranform, Medical information systems, Ultrasonic imaging, Ultrasonic variables measurement

I. INTRODUCTION ZOOSPERMIA is one of the causes of male infertility. It is the symptom that there is a complete absence of sperm in

the ejaculate. Fifteen-20% of infertile men turn out to be azoospermia. Azoospermia is caused by two problems; a production problem which is called non-obstructive azoospermia (NOA) and a deliver problem which is called obstructive azoospermia (OA). The development of intracytoplasmic sperm injection (ICSI) opened a new era in the field of assisted reproduction and revolutionized the assisted reproductive technique protocols of couples with male factor infertility [1]. Fertilization and pregnancies can be obtained with spermatozoa recovered not only from the ejaculate, but seminiferous tubules. The ideal technique of sperm extraction from testicle for NOA would be minimally invasive and avoid destruction of testicular function without compromising the chance of retrieval adequate numbers of spermatozoa to perform ISCI. Schlegel et

Manuscript received May 31, 2012. Y. Takashima is with Graduate School of Engineering, University of Hyogo,

Hyogo, Japan, and Himeji Initiative in Computational Medical and Health Technology, Hyogo, Japan (corresponding author to provide phone: +81-79-267-4986; e-mail: [email protected]).

T. Ishikawa is with Ishikawa hospital, Hyogo, Japan. K. Kuramoto, Syoji Kobashi and Y. Hata are with Graduate School of

Engineering, University of Hyogo, Hyogo, WPI Immunology Frontier Research Center, Osaka University, Osaka, and Himeji Initiative in Computational Medical and Health Technology, Hyogo, Japan.

al. have successfully developed the techniques with the assistance of an operating microscope as Micro-TESE [2]. TESE is a method of remove the seminiferous tubules in the testicle, and looking for the sperm in the seminiferous tubules. The possibility to including the sperm of larger seminiferous tubules is bigger than sclerotic seminiferous tubules. Some azoospermia patients have normal seminiferous tubules (Diameter: 0.250mm~0.300mm) and sclerotic seminiferous tubules (Diameter: 0.150mm). But there are the patients whose all tubules are small (Diameter: <0.150mm). The sperm cannot be collected if there are no large tubules in the testicle [3], [4]. Whether the sperm is able to be collected or impossibility is not understood if not operating. Additionally, the operation is costly and physically large burden for patients. Therefore, we need a system that noninvasively measures the thickness of the tubule testicular in the testicle before micro-TESE procedure. [5], [6].

Ultrasonic system is known as a kind of the medical devices for diagnosing human body. By using this device, several methods have been proposed to diagnose non-invasively the inside of the abdomen and blood [7]. Moreover, the ultrasonic quantitatively diagnoses inside of body with low cost and short time.

In our previous study [8], we have proposed a method to evaluate a rate of large line of a phantom by a 5.0MHz ultrasonic single probe. The phantom consists of large and small lines in total 24 in rubber tube. The previous method evaluated the rate of large line only. This study cannot evaluate the position of large line and small line.

In this study, we propose a seminiferous tubules evaluation system with an ultrasonic single probe. We can non-invasively evaluate it by frequency analysis. We employ two different diameter nylon fishing lines as a small tubule and large tubule. We calculate the frequency spectrum of the acquisition data by short-time Fourier transform. From the difference of frequency spectrum of the large and small lines, we sort these lines. We evaluate distance between the ultrasonic probe and reflectors. By using frequency value and the distance value, we make frequency map to visualize positions of large lines and that of small lines. As the results, the frequency map clearly showed positions of large and small lines. This experimental result suggested the high possibility to determine the larger diameter seminiferous tubule under the assumption that the characteristics of human testicular tubule are similar to that of the fishing line.

Ultrasonic Evaluation of Seminiferous Tubules by Frequency Map

Yuya Takashima, Student Member, IEEE, Tomomoto Ishikawa, Syoji Kobashi, Member, IEEE, Kei Kuramoto, Member, IEEE, Yutaka Hata, Fellow, IEEE

A

2012 Fifth International Conference on Emerging Trends in Engineering and Technology

978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.28

7


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.28

7

II. PRELIMINARIES

A. Phantom In our experiment, we employ two different diameter nylon

fishing lines as shown in Fig. 1. We employ these lines instead of seminiferous tubules. The diameter of the small line is 0.090 mm and that of the large line is 0.285 mm. Here, it is noted that diameter of the healthy seminiferous tubules ranged from 0.25 to 0.30 mm and that of the unhealthy tubules is smaller than 0.15 mm. Fig. 2 shows structure of phantoms we acquire in this study. In Fig.2 (a), it is put in 24 large rolled lines (25cm) and 24 small rolled lines in the rubber tube separately, and put in water. In Fig.2 (b), it is put in 24 large rolled lines (25cm) in mixed and water in the rubber tube.

B. Ultrasonic Data Acquisition System Fig. 3 shows an ultrasonic single probe. In this study, we

employ 5.0MHz ultrasonic probe. Our ultrasonic data acquisition system is shown in Fig. 4. Fig.

5 shows an example of the data acquisition situation. Sampling interval of the system is 20 ns. The ultrasonic waveform data are provided to a personal computer through the oscilloscope.

III. EVALUATION OF POSITION OF LARGE LINE Our method consists of six steps. Firstly, we acquire the

waveforms by the ultrasonic probe. We extract echo of the nylon lines. Secondly, we calculate frequency spectrum of waveforms by short-time Fourier transform. Thirdly, we calculate distance between probe and reflector. Fourthly, we reduce the noises by applying linear prediction for frequency spectrum. Fifthly, we calculate peak frequency value. Finally, we make frequency map from peak frequency value and distance value.

A. Frequency Analysis The system evaluates the diameter of the line by the frequency of ultrasonic waveform. A frequency of vibrating lines is determined by (1).

12

Tfl �

� (1)

Here, the notation f denotes frequency, l denotes length of line, T denotes tension of line, and � denotes linear density. In this research, we assume that l and T are constant. Thus, the frequencies are inversely proportional to the liner density as shown in (2).

1f�

� (2)

The density � is represented with a diameter, �, of line, as shown in (3).

Fig. 1. Large fishing nylon line and small fishing nylon line.

(a) Separated line phantom

(b)Mixed line phantom

Fig. 2. Phantom that put in lines and 15cm3 water.

Fig. 3. Appearance of 5.0 MHz ultrasonic single probe.

Fig. 4. Ultrasonic data acquisition system.

Fig. 5. Acquisition method.

2cm

Small Line

Large Line

Rubber Tube

2cm

Large Lineand

Small LineAcrylic board

Phantom

Plasticcontainer

Probe

88

2

4M�� (3)

Here, the notation M denotes degree of density. In this research we employ nylon line instead of seminiferous tubule for all phantoms. Therefore, M can be assumed as constant for all lines. By the same reason, M of seminiferous tubules are assumed as constant, Thus, relation (4) can be derived from (3).

2� �� (4)

From (2) and (4), frequency of a vibrating line depends on diameter of line (5).

1f�

� (5)

From this relation, we identify large line from small line by difference of the vibration frequency. Frequency of the large line is smaller than that of the small line. Fig. 6 shows the result that we examine frequency spectrum of 7 types lines. Fig. 7 is plotted with frequency f as vertical axis, and 1/� as horizontal axis. As shown in Fig. 7, there is a proportional relation between f and 1/�. Therefore we use this relation to evaluate position of large line.

B. Evaluation of Large line From Fig.7, we estimate large line from frequency spectrum.

In this section, we evaluate position of the large line. Ultrasonic reflection waveform has received time t series data of voltage. Here, received time t means the interval time between shots and receives echo from reflector at probe. Therefore distance d between the probe and the reflector is represented with t as shown in (6).

2

t Vd �� (6)

Here, the notation V denotes sonic velocity in the water. In this study we setup V condition as 1500 (m/s).

To calculate frequency spectrum of local sections of a waveform as it changes over time t, we employ short-time Fourier transform. From center t value, we calculate d value by (6) as a section’s distance from probe. We make frequency-map (F-map) from frequency spectrum of each section and its d value. This map is made with d as vertical axis, f as horizontal axis, and normalized power spectral density function (PSD) of frequency spectrum as 100 shades of gray brightness value. Fig. 8 shows an example of F-map which made by wave form obtained Fig. 2 phantom. As shown in this figure, where distance is shorter than about 2 (cm), the normalized PSD value of about 2 (MHz) (large line’s frequency) is large. This result means in this section large line is existing. On the other hand, where distance is about 2 to 3 (cm), the normalized PSD is large at about 5 (MHz) (small line’s frequency). This result means in this section small line is existing. In near section from probe, frequency spectrum was influenced by input voltage pulse.

C. Reducing Noises of Frequency Spectrum We reduce the noises by applying linear prediction for

frequency spectrum. We predict normalized PSD change of frequency by AR model [9]-[11]. The AR model is defined by (7).

1

( ) ( ) ( ) ( )p

iy f a i x f i u f

�

� � (7)

Fig. 6. Frequency spectrum of 7 lines.

Fig. 7. Calculation of Amplitude.

Fig. 8. F-map.

0

5

10

15

20

25

1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6

PSD

(dB

)

f (MHz)

90(�m)

117(�m)

155(�m)

205(�m)

235(�m)

260(�m)

285(�m)

y = 415.43x + 0.2976

0.0

1.0

2.0

3.0

4.0

5.0

6.0

0 0.002 0.004 0.006 0.008 0.01 0.012

f (M

Hz)

1/� (/�m)

99

Here, the notation y(f) denotes a predicted value of frequency f, x(f) does the normalized PSD value of f, a(i) does AR parameter, u(f) does white noise, and p does the order. AR parameter, a(i), is obtained from Yule-Walker equation defined by (8).

(0) (1) ( 1) (1) (1)(1) (0) ( 2) (2) (2)

( 1) ( 2) (0) ( ) ( )

R R R p a RR R R p a R

R p R p R a p R p

� � � � ��

� ��

��

� � � � � ��

(8)

Here, the notation R(i) denotes auto covariance function. In this method, we set the p as 30. Fig. 9 and Fig. 10 show the examples. In this result, noises were reduced. Fig. 11 shows

F-map make frequency spectrum applied linear prediction. As shown in Fig. 11, noises were reduced.

D. High Frequency and Low Frequency Maps We make low frequency map (LF-map) that shows position

of large line, and make high frequency map (HF-map) that shows position of small line to show position of large line and small line clearly. These maps are made with d as vertical axis, and peak frequency value fL and fH as 100 shades of gray brightness value. Here, notation fL denotes the lowest peak frequency and fH denotes the highest peak frequency. We make LF-map with fL value and make HF-map with fH value. We calculate peak frequency from frequency spectrum that applied linear prediction. On the frequency spectrum applied by linear prediction, we consider spectrum that the normalized PSD exceed a threshold value th in succession to be one part of frequency spectrum. In this method, we set the th as 0.05. In the one part of frequency, we define frequency that has the highest normalized PSD as the peak frequency. We cut out very low peak frequency on the supposition that spectrum of noise. We set the threshold as 0.4(MHz). We define the lowest peak frequency as the fL and the highest peak frequency as the fH. We show the example in Fig . 12. Fig.12 (a) shows the HF-map, and Fig. 12 (b) shows the LF-map obtained from separated line phantom (shown in Fig. 2 (a)). As shown in Fig. 12 high frequency exists in larger distance, and low frequency exists in short distance. This means, small lines are in long distance, and large lines are in smaller distance.

IV. EXPERIMENTAL RESULTS We employed nine phantoms. Two phantoms were made

from 24 large lines, 24 small lines and 15cm3 water in the rubber tube (Fig. 2 (a)). As shown in Fig. 2 (a) large lines and small lines were separated. Seven phantoms were made from large lines and small lines in total 24 rolled lines and 15cm3 water in the rubber tube as shown in Fig. 2 (b). We acquired ultrasonic wave forms. To compare our result with the other result, we show the result that employ amplitude value as brightness value (A-map). Amplitude value is obtained by Ref. [12].

Figures 13 to 17 show the result. In these figures, (a) shows

Fig. 9. Frequency spectrum.

Fig. 10. Normalized frequency spectrum applied linear prediction.

Fig. 11. F-map applied linear prediction.

(a)HF-map (a)LF-map

Fig. 12. HF-map and LF-map.

0.0

0.1

0.2

0.3

0.4

0.0 2.0 4.0 6.0 8.0

PSD

(dB

)

Frequency (MHz)

0.0

0.2

0.4

0.6

0.8

1.0

0.0 2.0 4.0 6.0 8.0

norm

lized

PSD

Frequency (MHz)

1010

F-map, (b) shows HF-map, (c) shows LF-map and (d) shows A-map. Fig. 13 shows the result obtained by separated line phantom which put large line part in probe face. Fig. 14 shows the result obtained by separated line phantom which put the small line part in probe face. Fig. 15 shows the result obtained by small lines only. Fig. 16 shows the results of large lines only. Fig. 17 shows the result obtained by 12 large lines and 12 small

lines. As shown in Fig. 13 high frequency exists in longer distance,

and low frequency exists in shorter distance. As shown in Fig. 14 high frequency exists in shorter distance, and low frequency exists in longer distance. To compare with A-map, HF-map shows position of small lines clearly, LF-map shows position of large lines clearly.

(a)F-map (a)HF-map (a)LF-map (a)A-map

Fig. 13. Results obtained by separated phantom (set up large line face).

(a)F-map (b)HF-map (c)LF-map (d)A-map

Fig. 14. Results obtained by separated phantom (set up small line face).


Fig. 15. Results obtained by small line only.


Fig. 16. Results obtained by large line only.

1111

Fig. 15 only shows high frequency, and Fig. 16 only shows low frequency. Fig. 17 shows high frequency and low frequency at the same distance. To compare with A-map, HF-map is confirmable exist of the small line, and LF-map is confirmable exist of the large line.

V. CONCLUSION In our study, we proposed a seminiferous tubules evaluation

system with a 5.0MHz ultrasonic single probe. Our proposed method understands approximate diameter of linear material and its distance from ultrasonic probe by using short-time Fourier transform. This method used here works more exactly than B-mode ultrasonic image.

In previous study Ref. [8], [12], from relation of diameter and frequency, we evaluate only rate of large line by amplitude value which applied band pass filters as an evaluate system before surgery. In this study, from same relation, we evaluate existing and distance of large line as a surgery support system.

In the future, we will do experiment with closer to the living body.

ACKNOWLEDGMENT This work was supported by a Hyogo Science and

Technology Association.

REFERENCES [1] G. Palermo, H. Joris, P. Devroey, A. C. Van Steirteghem, “Pregnancies

after intracytoplasmic injection of single spermatozoon into an oocyte,” Lancet 1992 vol.340, pp. 17–18.

[2] P. N. Schlegel, P. S. Li, “Microdissection TESE: spermatozoa retrieval in non-obstructive azoospermia,” Hum Reprod Update 1998 vol.4, pp. 439.

[3] T. Ishikawa, R. Nose, K. Yamaguchi, K. Chiba, and M. Fujisawa, “Learning curves of microdissection testicular sperm extraction for non-obstructive azoospermia,” Fertil Steril in press, pp. 1008-11, August 2010.

[4] T. Ishikawa, and M. Fujisawa, “Microdissection testicular sperm extraction for non-obstructive azoospermia: The assessment of serum hormone levels before and after procedure.” Japanese Journal of Reproductive Endocrinology, vol. 14, pp.15-20, 2009.

[5] A. Ramkumar, A. Lal, D. A. Paduch, and P. N. Schlegel,”AN Ultrasonically Actuated Silicon-Microprobe-Based Testicular Tubule Assay,” IEEE Transactions on biomedical engineering, vol. 56, pp. 2666-74, November 2009.

[6] J. L. Giffin, S.E.Franks, J. R. Rodariguez-Sosa, A. Hahnel, and P. M. Bartlewski, “A Study of Morphological and Haemodynamic

Determinants of Testicular Echotexture Characteristics in the Ram,” Exp Biol Med,pp. 794-801,2009 July.

[7] M. Nakamura, Y.T. Kitamura, T. Yanagida, S. Kobashi, K. Kuramoto, and Y. Hata, "Free placement trans-skull doppler system with 1.0MHz array ultrasonic probe," Proc. of 2010 IEEE Int. Conf. on Systems, Man and Cybernetics, pp. 1370-1374, 2010.

[8] Y. Takashima, K. Kuramoto, S. Kobashi, Y. Hata, T. Ishikawa, “Ultrasonic Thickness Evaluation of Seminiferous Tubule by Fuzzy Inference” Proc. of 2012 IEEE Int. Conf. on Fuzzy Systems,2012.

[9] J. Wang and T. Zhang, “Degradation prediction method by use of autoregressive algorithm,” IEEE ICIT, pp. 1-6, April 2008.

[10] F. M. Bennett, D. J. Christini, H. Ahmed, K. Lutchen, J. M. Hausdorff, N. Oriol, “Time series modeling of heart rate dynamics,” Computers in Cardiology, p.273, September 1993.

[11] W. Gersch and T. Brotherton, “AR model prediction of time series with trends and seasonalities: A contrast with Box-Jenkins modeling,” Decision and Control including the Symposium on Adaptive Processes, vol 19, p.988, December 1980.

[12] Y. Takashima, K. Kuramoto, S. Kobashi, Y. Hata, T. Ishikawa, “A Testicular Tubule Evaluation Method by Ultrasonic Array Probe” Proc. of 2011 IEEE Int. Conf. on Fuzzy Systems, pp. 1013-1016, 2011.


Fig. 17. Results obtained by mixed line phantom.

1212

A Visual Inspection System for Prescription Drugs

Takaaki MURAI, Masakazu MORIMOTO, Kensaku FUJII Himeji Initiative in Computational Medical and Health Technology

University of Hyogo Himeji, JAPAN

[email protected]

Abstract—To inspect prescription drugs with one-dose package (ODP), we propose an automated inspection system which based on computer vision. In the proposed system, we capture both-side of drugs at once and apply hierarchical identification. In this paper, we report several results of inspection experiments, which distinguish about a hundred kinds of naked drugs. As a result, we have achieved sufficient recognition rate and discuss its robustness.

Keywords- automated inspection; computer vision; one-dose package; log-polar transform; template matching

I. INTRODUCTION Recently, many dispensing pharmacies adopt one-dose

package (ODP) for prescribing drugs (figure 1). Most of drugs are prescribed with a press through package (PTP), but it sometimes causes miss-swallowing of their wrapping package. Therefore, to prevent any incorrect drug ingestion, a dispenser unwraps each drug and repackages them for each person according to dosing instruction. By this ODP, patient only needs to unpack and swallow single package at each time. It also prevents them from forgetting to take medicine.

There exists automated drug repackaging machine for ODP, it checks weight of package to confirm correct numbers of drugs were packed or not. However, final inspection still done by skilled person and it puts a heavy strain on dispensers. To solve this problem, we produce an automated ODP inspection system based on computer vision technology. There are many studies about product inspection of drugs in factory line [1], [2], but they are overmuch for pharmacies. So, our system simply consists of camera, PC and several instruments.

To identify prescription drugs, we can use their appearance information: size, shape, color and identifying mark. However, most of drugs have similar appearance, whitish color and circular shape, but some has peculiar one. So, in proposed system we first apply coarse classification to narrow down the candidates by basic appearance information, and then make the final decision by referencing identifying mark.

There are numerous variations of the identifying mark of drugs, ID characters with several fonts and alignments, weight notation number and emblem of company. Moreover, they were printed or engraved on single or both sides of drug. Therefore, we capture both-side images at once and apply template matching for verification. Here, we have two issues to resolve, one is randomness of drug direction and the other is appearance instability of engraved mark.

In our previous study [3], we focus on inspecting drug

tablets. In this paper, we report inspection results for both tablet and capsule drugs, which distinguish 100 kinds of naked drug tablets and 10 kinds of naked drug capsules. At this time we did not packing the drugs for one-dose, so effect of reflecting light from wrapping film and separation of contacted drugs are ignored.

Figure 1. One-dose package.

(a) Diagram of capturing unit

(b) Appearance of capturing unit

Figure 2. Drug capturing unit.


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.29

13


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.29

13

II. DRUG CAPTUREING UNIT To capture both sides of drugs at once, we can use two

cameras. However, it will raise the cost of hardware and cause calibration difficulty between two cameras. So, in this paper, we use single camera with mirror unit. Figure 2(a) shows a diagram of our capturing unit and its appearance is shown in Fig. 2(b). It consist of two mirrors which are right angles to each other, transparent board to put on drugs, diffused panel light at both sides and capturing camera. By this unit, we can capture both-side images of drugs almost same condition.

Figure 3 shows an example of captured drugs. We can easily find correspondence between upper and bottom sides of the same drug, because it locates symmetrically in a picture. For the following experiments, we prepare 100 kinds of drug tablets and 10 kinds of drug capsules. We take 10 pictures for each drug tablets and 25 pictures for each drug capsules with different position and different direction. Figure 4 shows some examples of captured drug images.

III. PRESCRIPTION DRUG IDENTIFICATION To identify prescription drugs, identifying mark is an

important clue. However, there are numerous variations of the identifying mark: printed or engraved, character or emblem, type of font and their alignments. Furthermore, there exist several fluctuations of ID mark appearance, such as lighting condition, distortion due to capsule shape, rolling of capsules and rotation of tablet direction.

To compensate these fluctuations, first we distinguish between tablets and capsules. They have different fluctuation tendency, we can simplify ID mark extraction method by adapting for tablets and capsules, respectively.

A. Distinguish tablets and capsules When a captured drug doesn’t have a circular form, it might

be a drug capsule. To distinguish capsule or not, we focus on curvature of long side of the elliptical drug, because most of elliptical tablets have arc at longer side. To align elliptical drug images, we detect both end of drug region as circles by using Hough transform [4]. The line which connects centers of detected circles is long axis of the ellipse. Then we align the capsule images by using this axis.

Figure 5 shows shape alignment result of three kinds of elliptical tablets and one typical capsule. As shown in this figure, most of elliptical tablet have enough thickness of arc at longer side. Only one out of 10 elliptical tablets (Fig. 5(c)) has a shape very similar to capsules. For such tablets, we have to register it both tablets and capsules, and deal it as a capsule at inspection stage.

B. Tablet Identification To realize tablet identification, we apply mark extraction

and rotation compensation before template-matching. We first extract identifying marks from input images to avoid lighting instability. Here, we apply Difference-of-Gaussian (DOG) to extract identifying marks. In DOG, the image �(�, �) is first smoothed by convolution with Gaussian kernel of certain width σ� by following equation,

Figure 3. A captured image.

(a) drug tablets

(b) drug capsules

Figure 4. Captured drug images.

1414

to get smoothed image ��(�, �)

With a different width σ� , a second smoothed image can be obtained:

The difference of these two Gaussian smoothed images enhances edge and discards gradual luminance fluctuation of input image. Then we remove outline and small noise from output of DOG to extract identifying mark only. Figure 6 shows mark extraction results for printed and engraved drugs. As shown in this figure, we can treat printed mark and engraved mark identifiable.

To realize image registration between two randomly rotated images, we use log-polar transform [5]. Figure 7 shows examples of the transformed image. This is a coordinate transformation and has two notable features. One is that the difference of rotation angle between input images can be represented by vertical linear shift of transformed image. Another is that scale difference between two images can be represented by horizontal shift of transformed image. For general image registration, selection of center point for transformation is a considerable issue. However, for drug identification, center point can be uniquely determined by center of drug region. By the log-polar image registration, we can achieve rotation compensated query input image.

(a) elliptical tablet A

(b) elliptical tablet B

(c) elliptical tablet C

(d) drug capsule

Figure 5. A captured image.

(�, �, �) = 1�2��

�� − �� + ��2��

� (1)

��(x, y) = (�, �, �) ∗ �(�, �). (2)

��(x, y) = (�, �, �) ∗ �(�, �). (3)

(a) Input grayscale images (left: Printed, right: Engraved)

(b) Output of Difference-of-Gaussian

(c) Denoising and contour elimination Figure 6 Preprocessing of input image

(a) Image in database (b) Query input image

(c) Log-polar transformed images.

Figure 7 Log-polar transformations to adjust rotational uncertainty.

1515

C. Capsule Identification Most of medical capsules have cylindrical shape. This

shape causes two problems for visual inspection. One problem is distortion of ID marks which located around capsule edge in captured image. Another problem is variation of ID mark appearance. Fig. 8(a) shows an example of variation caused by rolling. Furthermore, appearance of ID mark is unstable because medical capsule is consisted of two parts, body and cap (Fig. 8(b)), and their combination differs each other.

Capsules’ ID marks are distorted because capsules have cylindrical shape. To solve this problem, correct distortion with extending the capsules image (Fig. 9). The intensity of the distortion can be presumed from the width of the mask image and diameter of the capsule.

(a) Appearance variation by rolling

(b) Appearance variation by parts

combination

Figure 8. Variation of ID mark appearance

(a) Distorted image

(b) Corrected image

Figure 9. Collecting distortion

(a) ID image of the cap

(b) ID image of the body

Figure 10. Extracted ID images

(a) Training image of the cap

(b) Training image of the body

Figure 11. Training image of the capsule

A Capsule’s training images are required to contain all- around ID image of the capsule’s body and cap (Fig. 10). So,

we take a number of capsule images with rolling the capsule slightly at each time. Then connect ID images and make whole around ID image as shown in Fig. 11.

IV. PRELIMINARY EXPERIMENTAL RESULTS

A. Tablet recognition results To evaluate our proposed system, we have recognized 100

kinds of naked drugs. Here, we register two out of ten images of a drug as database and rest eight images are used for recognition test. As a result, we could distinguish whole 100 type’s 800 drug images correctly. However, average pharmacies store generally turn over 500 kinds of drugs, so we have to estimate the performance for realistic situation.

To estimate its robustness, we compare 1st correct candidate and 2nd incorrect candidate for each query input. Here, we show normalized correlation factor of 2nd incorrect candidate, which adjust the correlation factor of correct 1st candidate becomes to 1.0. As a result, normalized correlation factors are less than 0.9 for 99% of input query images. This result implies that our drug recognition system have sufficient ability.

The consideration for the rest 1% confusing query images, which exceed 0.9 normalized correlation factors, will provide a clue for further improvement. Figure 12 and 13 show drug pairs which have the highest normalized correlation factor for printed and engraved drugs, 0.966 and 0.873, respectively.

From these figures, it is clear that the most confusingly drug pair is produced by identical pharmaceutical company. They have the same company logo and similar appearance, only have different product code. Therefore, to realize accurate inspection system, we have to list up the similar drug pairs at registration step, and prepare weight map for distinguishing those pairs.

In this experiment, we register orthogonal two images as database, because it had not work well when we register only one image for each drug. This is caused by instability of mark extraction due to the lighting direction for engraved drugs. To confirm effect of the instability, we analyze cross-correlations between different angle images shown in figure 14. Here, for comparison, we also get cross-correlation factors between the similar drug pair shown in figure 13. From table I, we can see that the correlation factor between same drugs with different direction below the value of similar pairs. These result implies that our system have room for improvement in lighting environment and mark extraction procedure.

TABLE I. CROSS-CORRELATION FACTORS BETWEEN TWO KINDS OF DRUGS WITH DIFFERENT ANGLES

Drug Type “AK401” “AK421”

A B A B

“AK 401”

A 1 0.49 0.67 0.45

B 0.49 1 0.46 0.55

“AK 421”

A 0.64 0.45 1 0.59

B 0.45 0.56 0.43 1

1616

B. Capsule recognition results When we recognize capsules only by ID marks,

recognition rate becomes 97.8%. In this case, the misrecognized capsules have similar ID mark, which often have only one digit difference between each ID. In order to reduce such misrecognition, we have to use color information to distinguish similar ID images.

By adding color information, we could identify 223 of 225 images by our proposed method, and its recognition rate becomes 99.1%. Fig. 15 shows a graph of recognition rate of this experiment and the ID image recognition case. This figure shows that the color information improves recognition rate and correct candidate becomes higher rank. The misrecognition images of this experiment are shown in Fig. 16. From this figure, we can find that the misrecognition is caused by capsules which have similarity for both color and ID mark.

We also examined the number of candidates which are inspected by the color recognition. We found that the color recognition was carried out only when a similar capsule exists and the accurate of the ID image is low. Fig. 17 shows a histogram of the number of candidates and the number of inspection images. From this experiment, we found that the color recognition can improve recognition rate. The color recognition is expected that the method will be used for rough classification before the ID image recognition. However, we must think out plan for reducing misrecognition caused by the color recognition.

Finally, we could not recognize two capsules in the experiments. These capsules are made by same pharmaceutical company for the same purpose. The capsules differ in the amount of components. It is rare case that two capsules have very similar colors and ID marks. Capsules have a different design for each usage (Fig. 18) even if they are made by same company because pharmaceutical companies design their products to be easily distinguished from appearance. Even if capsules for same purpose have similar ID marks, most of them have different colors and size (Fig. 19). From those reasons, our proposed method can prevent prescription of wrong purpose drugs because the method can easily distinguish two different designed capsules. However we must consider the method for recognizing very similar capsules to recognize various capsules and to improve recognition rate.

Figure 15 Comparing recognition rate

96

97

98

99

100

1 2 3 4 5 6 7 8 9

Rec

ogni

tion

rate

[%]

Candidate rank

Print and color inspectionOnly print inspection

Figure 12 A printed drug pair which has normalized correlation 0.966

Figure 13 A engraved drug pair which has normalized correlation 0.873

(a) “AK401”-A (b) “AK401”-B

(c) “AK421”-A (d) “AK421”-B

Figure 14 Extracted mark images for two kinds of engrved drugs with different angles.

1717

(a) Inspection images 1

(b) First candidate capsule 1

(c) Inspection images 2

(d) First candidate capsule 2

Figure 16. Examples of misrecognition images

Figure 17. The number of candidates of color recognition

Figure 18. Capsules for different purposes

Figure 19. Variation of the amount of components

V. CONCLUSION In this paper, we report several results of preliminary

experiments for drug inspection system which targets on ODP prescription drug inspection. Here we targeted both tablets and capsules simultaneously. As a result, we could distinguish whole 100 type’s 800 drug tablets and achieve 97.8% of capsule inspection accuracy.

At this time we didn’t packing the drugs for one-dose, so effect of reflecting light from wrapping film and separation of contacted drugs are ignored. Therefore, in further study, we have to tackle these issues.

References [1] M. Varghese and C. Cetinkaya, “Noncontact photo-acoustic defect

detection in drug drugs,” Journal of Pharmaceutical Sciences, vol. 96, pp.2125-2133, 2007.

[2] Y. C. Shen and P. F. Taday, “Development and application of terahertz pulsed imaging for nondestructive inspection of pharmaceutical drug,” IEEE Journal of Selected Topics in Quantum Electronics, vol.14, no.2, 2008.

[3] M. Morimoto, K. Fujii, “A visual inspection system for drug tablets,” IEEE SMC2011, pp.1106-1110.

[4] H.K. Yuen, J. Princen, J. Illingworth, J. Kittler, "A comparative study of Hough Transform methods for circle finding," BMVC 1989, pp.169-174.

[5] G. Wolberg and S. Zokai, “Robust image registration using log-polar transform,” Proceedings of IEEE International Conference on Image Processing (ICIP 2000), vol.1, pp.493-496, 2000.

30

5

0 0

3 1 0 0

0

5

10

15

20

25

30

35

2 3 4 5 6 7 8 9

Num

ber

of in

spec

tion

imag

es

Number of candidates

1818

On Waiting-time Estimation for Outpatients at Department of Ophthalmology

Naotake Kamiura, Ayumu Saitoh, Teijiro Isokawa,

and Nobuyuki Matsui

Dept. of Electrical Eng. and Computer Sciences

Graduate School of Engineering, University of Hyogo

Himeji, Japan

{kamiura, saitoh, isokawa, matsui}@eng.u-hyogo.ac.jp

Hitoshi Tabuchi

Department of Ophthalmology

Tsukazaki Hospital

Himeji, Japan

[email protected]

Abstract—In this paper, waiting-time estimation is presented

for outpatients visiting department of ophthalmology. It

determines the number of virtual outpatients according to the

probability density functions of the gamma distribution.

Assignments of arrival time intervals for such outpatients

depend on random numbers generated from the exponential

distribution. It assumes that outpatients undergo all

examinations within the waiting time. The examination order

is determined by six rules established by the surveys at the

hospital. Once a target outpatient specifies a day of the week

and arrival time of the outpatient desiring, disease name, and

name of ophthalmologist at the time of making an appointment,

the proposed system calculates the waiting time for the

outpatient. Experimental results establish that the difference

between estimated waiting time and actual waiting time is

acceptable to outpatients.

Keywords- department of ophthalmology; examination-time

estimation; outpatients; waiting-time estimation

I. INTRODUCTION

In Japan, long waiting time certainly decreases hospital customer satisfaction rating [1]. In addition, it has been considered to be one of the major reasons that prevent sick persons from going to hospitals. The condition of such persons keeps on worsening while they hesitate to go to hospitals, and the fatal damage tends to suddenly befall to them. A straightforward approach to overcome this issue is to save even short amount time in the waiting room. In [2], an event-driven network approach using queuing theory is presented to reduce the waiting time of patients. The dispatching rules are suggested based on patients’ expected visitation time and expected service time, and they are used to schedule the patients.

The other approach is to reduce the stress of outpatients by means of disclosure of information on waiting time. In [3], a waiting time prediction system for consultations in hospitals is developed. It obtains arrival time, examination-start and finish time, and so forth, for every outpatient in real time by way of LAN. Based on such data, it updates the consultation order of outpatients and estimates the waiting time in real time. The estimated waiting time is announced to outpatients on a display. After the consultation on some day is complete, the outpatient generally makes an appointment for next consultation while discussing it with a medical doctor. Since the system in [3] requires real-time

data on the day when the outpatient consults the doctor, it is impossible to apply the system when making the appointment for next consultation.

In this paper, waiting-time estimation available in making appointments is presented for outpatients visiting the department of ophthalmology in a hospital. The proposed method considers the waiting time to be the difference between the time point when an outpatient begins to consult an ophthalmologist and arrival time of the outpatient, and assumes that all examinations imposed on the outpatient must be complete within the waiting time [4]. It determines the number of outpatients virtually visiting the department during some time period according to the probability density functions of the gamma distribution. It then assigns arrival time intervals for such virtual outpatients according to the random numbers based on the exponential distribution. The gamma distribution and exponential distribution are defined according to statistics provided from the hospital. Six rules determine the order of examinations imposed on outpatients contracting major diseases, and are used to estimate the time consumed by the examinations. An outpatient can obtain the waiting-time information by giving the choice associated with a day of the week and arrival time of the outpatient desiring, disease name, and name of ophthalmologist to the proposed system. It is experimentally revealed that the difference between the waiting time estimated by the proposed method and the actual time calculated from the record in 2011 is within the allowable range.

II. PRERIMINARIES

In this paper, exponential distribution and gamma distribution are employed to model changes in the number of outpatients through time. They are fundamental functions in queuing theory [5].

The exponential distribution is generally considered to be suitable in expressing the distribution associated with time at which an event occurs. The probability density function is given by

f (x) =�e��x if x � 0,

0 otherwise.

� � �

(1)

When this function is applied to a business establishment, the

parameter � means the number of customers visiting it per


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.23

24


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.23

24

unit time on the assumption that two or more customers do not come simultaneously. Note that x is the time interval of

customers arriving. 1/� then equals the mean interval

associated with arrival time points. The proposed method generates random numbers based

on the exponential distribution, and employs them to determine the time interval of outpatients arriving. The numbers are uniformly distributed in the range [0, 1) . Let y

denote the exponential random number obtained by uniform random number u, which belongs to the range [0, 1) and is

generated according to inverse function rule. It is given by

y = �1

�loge 1� u( ). (2)

The value of y then corresponds to the time interval of outpatients arriving.

In this paper, it is assumed that the distribution associated with the number of outpatients visiting the department of ophthalmology at some time point depends on the gamma distribution with parameter k. The probability density function of the gamma distribution is as follows.

f (xi) =

(�k)k

�(k)xik�1e��kxi if 0 � xi <�,

0 if xi < 0,

�

�

(3)

where k>0 and �>0. xi means the i-th time interval. An

interval equals ten minutes. The value of � equals the mean

number of outpatients visiting the department per ten

minutes period. The gamma function is denoted by �(k).

The value of f(xi) denotes the number of outpatients visiting the department for the i-th time interval. The proposed method prepares f(xi)’s that change depending not only on the day of the week but on the morning and the afternoon.

III. WAITING TIME EXPECTION AVAILABLE IN MAKING

APPOINTMENTS

A. Assignment of Time Intervals of Outpatients Arriving

It is defined for an outpatient that the waiting time is equal to the difference between the time point when the outpatient arrives at the department of ophthalmology and the time point when an ophthalmologist begins to make a diagnosis for the outpatient. It is also assumed that the outpatient usually undergoes some examinations before he/she consults the ophthalmologist. The proposed method determines examinations and the order of them for virtual outpatients, and simulates the passage of time required for the examinations and consultation imposed upon them. The simulation requires the following statistics and record: the mean number of outpatients on each day of the week, time intervals of outpatients arriving during each period of time on each day of the week, the distribution of arrival time for outpatients on each day of the week, diseases contracted by outpatients arriving on each day of the week, names of ophthalmologists working on each day of the week, names of

ophthalmologists being skilled at treating each disease, time required for consultation made by each ophthalmologist, and the consultation time for each disease. The proposed method uses the above accumulated in 2011 at Tsukazaki hospital, Japan.

Let us assume that it is necessary to estimate the waiting

time for outpatient � with disease � at time denoted by � �

hours (i.e., �:�) on day �, where 8��17, 0��59 and � is a

member of the set of seven days, {Sunday, Monday, Tuesday, Wednesday, Thursday, Friday, Saturday}. The proposed method then uses one of the above statistics (i.e., the mean number of outpatients visiting the department of

ophthalmology on the target day same as �). In addition, it

assigns arrival time, disease name, and ophthalmologist

name to each of virtual outpatients excluding �, according to

the above statistics. Assignment of the disease name depends on the ratio of the number of outpatients contracting each disease compared to the total number of outpatients visiting the department on the target day. For the outpatients contracting some disease, the proposed method also calculates the ratio of the number of them consulting each ophthalmologist compared to the total number of outpatients on the target day. This calculation is made for every disease, and the ophthalmologist name is determined in accordance with the ratio calculated in such a way. Specifying the consultation time for each virtual outpatient accompanies assigning the disease name and the ophthalmologist name to him/her. In other words, the mean consultation time, which is consumed by each ophthalmologist on condition that an outpatient contracts each disease, is employed.

Fig. 1 depicts the approximate outline of changes in the number of outpatients arriving on a day of the week through time. It is safety considered that the change in the morning (or in the afternoon) is quite similar to the gamma distribution. This is why it is described that the number of outpatients arriving depends on the gamma distribution in Section II. Note that the number of outpatients visiting the department in the morning is generally larger than that in the afternoon as shown in Fig. 1. The proposed method prepares a probability density function for the morning and that for the afternoon, individually, on a day of the week. It divides office hours in a day into a large number of ten minutes periods, and determines the number of virtual outpatients arriving per ten minutes period according to the probability density functions.

The proposed method then assigns arrival time intervals to virtual outpatients visiting the department during every ten minutes as follows. For a given ten minutes period, a statistical time interval of virtual outpatients arriving is first calculated. It is equal to the mean interval obtained from accumulated data associated with the above period in 2011. The proposed method next defines the exponential distribution for the above period by using the mean interval. Recall that we have the number of virtual outpatients arriving during the above period, based on the gamma distribution. The arrival time intervals to be assigned to such outpatients for the above period depend the random numbers based on the above exponential distribution.

2525

Figure 1.

Approximate outline of changes in outpatient number on a day.

It is impossible to assign arrival time intervals for a given ten minutes period in the above manner, if the mean interval associated with the period equals zero. This is due to the fact that the exponential distribution can never be defined by using the mean interval. The arrival time interval is then set to the value equal to ten divided by the virtual outpatient number obtained from the gamma distribution for the above period.

B. Examination Time

Examinations conducted at Tsukazaki hospital are tabulated in Table I. An examination group is defined for each of major diseases, and the major diseases are assigned to virtual outpatients as disease names. The examination group consists of usual examinations and prescribed examinations. The prescribed examinations are indicated by boldface in Table I, and orders from ophthalmologists are specially required for execution of them. In other word, the prescribed examinations are often cancelled by ophthalmologists. Visual acuity test, refraction test, tonometry, slit lamp examination, and mydriatic test are referred to as major examinations. One of the major examinations appears in every examination group at least. A set of visual acuity test, refraction test, tonometry, and axial length measurement is specially referred to as the principal examination set. The outpatients undergo pupil examination at the unit designed for visual acuity test, and hence the proposed method considers it to be part of visual acuity test to estimate the time consumed by them. The following three equipments are prepared for outpatients to undergo slit lamp examination: optical coherence tomography (OCT) for anterior eye segment, wavefront sensor, and dynamic contour tonometer (DCT).

Except for major examinations, the examinations in Table I are referred to as infrequent examinations. They are categorized as shown in Table II. A set of infrequent examinations includes a set of prescribed examinations. Tsukazaki hospital sets up equipments designed for checking the optic nerve, ophthalmoscopy, optical coherence tomography, and fundus angiography in the same room, and hence Category 1 consists of them. This also applies to

Category 2. The infrequent examinations excluded from Categories 1 and 2 appear in Category 3. The probabilities that ophthalmologists determine whether outpatients should undergo prescribed examinations in Table I are given in Table III. They are based on the survey completed at Tsukazaki hospital. Table III shows that both of the optical coherence tomography and fundus angiography are needed for outpatients contracting one of the diabetes, retinal disease, and uveitis at 11.5 percent. This also applies to a couple of mydriatic test and fundus angiography for outpatients contracting neuro-ophthalmologic disease. The proposed method assumes that virtual outpatients undergo prescribed examinations according to percentages in Table III.

Let us next discuss examination time. Mean consumption time per examination is tabulated in Table IV. The values in Table IV are also based on the survey. Each of the values associated with Categories 1, 2, and 3 is equal to mean time consumed by an examination belonging to each category. Since optical coherence tomography (OCT) for anterior eye segment, wavefront sensor, and dynamic contour tonometer (DCT) are not always used for every outpatient undergoing slit lamp examination, mean time consumed by them is separately given as Slit lamp examination 1, Slit lamp examination 2, and Slit lamp examination 3. One of the following combinations is executed as slit lamp examination: a combination of inspections using OCT and wavefront sensor, that using OCT, wavefront, and DCT, and that using DCT solely. The first, second and third combinations are referred to as CSL1, CSL2 and CSL3, respectively.

To make the calculation of total examination time for every virtual outpatient realistic, time consumed by some examination is changed as follows. A coefficient denoted by Cbase is defined, and the baseline time is calculated as the product of each value in Table IV and Cbase. The baseline time is then multiplied by 1+Cpersonal, and the resultant value is set to the examination time for some virtual outpatient, where Cpersonal is randomly determined for the virtual

outpatient from the range of �0.1 to +0.1. The proposed

method adds fifteen seconds to the above product of baseline time and 1+Cpersonal as time to walk from equipment to another equipment. In other words, moving time between the two examination equipments equals fifteen seconds.

For visual acuity test, refraction test, and visual field test, there exist outpatients that require extremely longer (or shorter) time than other outpatients undergoing the same examination. To overcome this problem, the exceptional mean time in Table V is employed. In addition, Table V shows the percentage of the number of outpatients undergoing one of the examinations with the above time compared to the total number of outpatients undergoing the same examination. Both of the mean time and percentages are also based on the survey. The proposed method randomly chooses some of virtual outpatients according to percentages in Table V, and assumes that the chosen outpatients undergo visual acuity test, refraction test, and visual field test with the baseline time equal to the product of the mean time in Table V and Cbase.

2626

C. Rules for Order of Examinations

The proposed method determines examination order in accordance with the following rules. They are established according to the survey completed at Tsukazaki hospital. Rule 1: If necessary, a virtual outpatient undergoes axial length measurement before refraction test and tonometry. Rule 2: The number of units designed for visual acuity test is five. If all of the five units are occupied, a virtual outpatient waits. Once one of them is available, the virtual outpatient undergoes visual acuity test using it. Rule 3: If all of the equipments are unavailable for axial length measurement and refraction test while at least one of

the visual acuity test units is available, a virtual outpatient undergoes visual acuity test before axial length measurement and refraction test. Rule 4: A virtual outpatient undergoes examinations belonging to the principal examination set before slit lamp examination and examinations of Category 2. The virtual outpatient then undergoes slit lamp examination and examinations of Category 2 as soon as one of them is available. If both of slit lamp examination and Category 2 are simultaneously available, the execution order for them is randomly determined. Rule 5: For CSL1 and CSL2 executed as slip lamp examination, the order of inspections is randomly

TABLE I. EXAMINATIONS CONDUCTED AT HOSPITAL

Group

number Major diseases Examinations

1 Cataract Visual acuity test, Refraction test, Tonometry, Slit lamp examination, Mydriatic test, Axial length measurement

2 Glaucoma Visual acuity test, Refraction test, Tonometry, Slit lamp examination, Mydriatic test, Checking the optic nerve,

Visual field test, Axial length measurement

3 Diabetes Visual acuity test, Refraction test, Tonometry, Slit lamp examination, Mydriatic test, Ophthalmoscopy, Optical

coherence tomography, Fundus angiography

4 Retinal disease Visual acuity test, Refraction test, Tonometry, Slit lamp examination, Mydriatic test, Ophthalmoscopy, Optical

coherence tomography, Fundus angiography

5 Corneal and/or

conjunctival disease Visual acuity test, Refraction test, Tonometry, Slit lamp examination, Lacrimal test

6 Uveitis Visual acuity test, Refraction test, Tonometry, Slit lamp examination, Mydriatic test, Ophthalmoscopy, Blood test,

Optical coherence tomography, Fundus angiography

7 Strabismus Visual acuity test, Refraction test, Eye movement examination

8 Neuro-ophthalmologic

disease

Visual acuity test, Refraction test, Tonometry, Pupil examination, Slit lamp examination, Eye movement

examination, Critical flicker fusion test, Visual field test, Mydriatic test, Fundus angiography

9 Lacrimal duct

disorder

Visual acuity test, Refraction test, Tonometry, Slit lamp examination, Dacryoscintigraphy or dacryoendoscopy or

dacryoendoscopy, Tear drainage test

10 Pediatric

ophthalmology Visual acuity test, Refraction test, Tonometry, Slit lamp examination, Mydriatic test, Ophthalmoscopy

11 Refractive error Visual acuity test, Refraction test, Tonometry, Slit lamp examination

12 Trauma Visual acuity test, Refraction test, Tonometry, Pupil examination, Eye movement examination, Mydriatic test,

Ophthalmoscopy, Orbit CT scan

13 Other symptoms Visual acuity test, Refraction test, Tonometry, Slit lamp examination

TABLE II. INFREQUENT EXAMINATIONS

Category number Infrequent examinations

1 Checking the optic nerve, Ophthalmoscopy, Optical coherence tomography, Fundus angiography

2 Visual field test, Eye movement examination

3 Axial length measurement, Lacrimal test, Blood test, Critical flicker fusion test, Tear drainage test,

Dacryoscintigraphy or dacryoendoscopy or dacryoendoscopy, Orbit CT scan

TABLE III. PROBABILITIES THAT OPHTHALMOLOGISTS DETERMINE WHETHER OUTPATIENTS SHOULD UNDERGO PRESCRIBED EXAMINATIONS

Group number Major disease Prescribed examinations�Probabilities of examination

execution (percent)

1 Cataract

2 Glaucoma Axial length measurement 15.0 %

3 Diabetes

4 Retinal disease

6 Uveitis

Optical coherence tomography,

Fundus angiography 11.5 %

5 Corneal and/or conjunctival

disease Lacrimal test 50.0 %

Visual field test 10.0 % 8 Neuro-ophthalmologic disease

Mydriatic test, Fundus angiography 11.5 %

9 Lacrimal duct disorder Tear drainage test 50.0 %

12 Trauma Orbit CT scan 50.0 %

2727

determined. Note that, for CSL3, the end of inspection using DCT means the end of slip lamp examination. Rule 6: If either mydriatic test or optical coherence tomography (OCT) is imposed on a virtual outpatient, he/she undergoes it at the end of an examination sequence. If both of them are imposed, the virtual outpatient undergoes mydriatic test and OCT in that order.

Examinations not mentioned in the rules are executed in the order of equipments designated for them being available.

IV. EXPERIMENTAL RESULTS

In the following, the actual outpatient that the proposed method estimates the waiting time for is referred to as the target outpatient. Users (e.g., the ophthalmologist and target outpatient) first give the following requests of the target outpatient to the proposed system: a day of the week and arrival time of the target outpatient desiring, disease name, and name of ophthalmologist in charge. On the other hand, the proposed system autonomously determines ID number, arrival time point, disease name, examinations to be imposed, the order of them, start time point and finish time point for each of them, examination time for each of them, name of ophthalmologist in charge of consultation, and consultation time for every virtual outpatient. Determining ID number depends on the order of arrival of a virtual outpatient. Recall that fifteen seconds are set to moving time between the two examination equipments. If two examinations are imposed on some outpatient, start time point of the latter examination for the outpatient is the latest time point when fifteen seconds pass after the former examination is complete for him/her and the equipment for the latter examination is available.

Let us assume that it is necessary to estimate the waiting

time for some target outpatient with disease � at time point

denoted by � � hours (i.e., �:�) on day �, where 8��17,

0��59, � is a member of the set of days of the week, and �

is one of the major diseases shown in Table I. The proposed

method then sets the starting point of simulation to � o’clock.

The simulation finishes at the start point of consultation for the target outpatient. The proposed method produces a list of arrival time, start and finish time points of imposed examinations, start time point of consultation and waiting time associated with the target outpatient. Recall that waiting time is equal to the difference between start time point of consultation and arrival time. In this paper, it is estimated subject to Cbase=0.8.

The proposed method is first evaluated in terms of the average waiting time for each day of the week. The ophthalmologist that the most outpatients consult in Tsukazaki hospital is checked for each day, according to the record in 2011. The disease contracted by the most outpatients consulting such ophthalmologist is also checked for each day. The arrival time is then assumed to be p

hundred hours (i.e., p o’clock) for each day where 9�p�16.

The proposed method estimates the waiting time on condition specified by the above ophthalmologist, disease and arrival time. The above is repeated ten times for every arrival time, and the average waiting time is acquired for each day. Table VI shows the simulation results. Note that the real average waiting time calculated from the record in 2011 on the above condition appears in Table VI. Table VI remains ophthalmologist name anonymous. It is established in [3] that outpatients tolerate the system performance if the difference between its output (i.e., waiting time estimated by the system) and the actuality (i.e., actual waiting time) is less than thirty minutes. All of the average results in Table VI are within this standard.

Let us next discuss the change in estimation with the passage of time. The arrival time of the target outpatient is set every thirty minutes from 8:30 until 16:00 as shown in Table VII. Note that time points 12:00 and 12:30 are excluded. In the record, the number of outpatients visiting the department of ophthalmology in the above hospital on Saturday is larger than that on any other day. Saturday is therefore chosen for the simulation. The ophthalmologist that the most outpatients consult on Saturday is also chosen. The number of aged outpatients tends to be larger than that of young outpatients, and cataract and glaucoma are common diseases for the aged outpatients as shown in Table VI. This is why outpatients contracting either cataract or glaucoma are explored in terms of the waiting time. The proposed method estimates the waiting time on the above condition, and the evaluation is repeated ten times for every arrival time. Table VII shows the averaged results. It also shows the real average waiting time calculated from the record on the same condition. The difference between the estimation and the actuality never exceeds thirty minutes for every arrival time. The proposed method thus makes it possible to lessen discontent from ophthalmologists and outpatients with waiting time.

V. CONCLUSION

In this paper, a method of estimating waiting time has been proposed for outpatients visiting department of ophthalmology. It is designed to be available at any time point when appointments are made. It determines the number of virtual outpatients per ten minutes period

TABLE IV. MEAN CONSUMPTION TIME PER EXAMINATION

Examinations Mean consumption time

per examination (sec.)

Tonometry 60

Refraction test 90

Visual acuity test 600

Slit lamp examination 1 (OCT

for anterior eye segment) 240

Slit lamp examination 2

(wavefront sensor) 280

Slit lamp examination 3 (DCT) 220

Mydriatic test 1200

An examination belonging to

Category 1 360


Category 2 240


Category 3 210

2828

according to the probability density functions of the gamma distribution. It then assigns arrival time intervals for virtual outpatients according to the random numbers based on the exponential distribution. Assuming that outpatients undergo all examinations within the waiting time, it establishes six rules associated with the order of the examinations. The rules are used to estimate the time consumed by the examinations. Once a target outpatient specifies the choice associated with a day of the week and arrival time of the outpatient desiring, disease name, and name of ophthalmologist, the proposed system calculates the waiting time required for the outpatient. Experimental average results associated with the difference between the estimation and the actuality are less than thirty minutes both for each day of the week and for every half hour. It is thus revealed that the proposed method is useful in relieving outpatients’ stresses caused by waiting time.

In future, the proposed method will be modified to improve the estimation accuracy.

REFERENCES

[1] M. Miyamoto, “Waiting time for outpatient services,” (in Japanese) Keizaikei, Quarterly Journal of Economics, The Society of

Economics, Kanto Gakuin University, vol. 226, pp.1-12, Jan. 2006.

[2] Y. Ji, Y. Yanagawa, and S. Miyazaki, “Reducing outpatient waiting times for hospital using queuing theory,” (in Japanese) Journal of

Japan Industrial Management Association, Vol. 60, No. 6, pp. 297-305, Feb. 2010.

[3] A. Ogawa, K. Maki, N. Esumi, H. Asahara, S. Kawauchi and Y.

Takeuchi, “Development of a waiting time prediction system for hospitals,” (in Japanese) IEEJ Transactions on Electronics,

Information and Systems, Vol. 129, Issue 7, pp. 1264-1268, July 2009.

[4] N. Kamiura, A. Saitoh, T. Isokawa, N. Matsui, and H. Tabuchi, “On classification of interview sheets for ophthalmic examinations using

self-organizing maps,” Proc. of the 17th International Conference on Artificial Life and Robotics (AROB 17th '12), Artificial Life and

Robotics, Jan. 2012, pp. 686-699.

[5] T. Honma, Queuing theory, (in Japanese) Rikohgakusha, Tokyo, 1966.

TABLE V. EXCEPTONAL MEAN TIME FOR VISUAL ACUITY TEST, REFRACTION TEST, AND VISUAL FIELD TEST

Group

number Major disease Examinations�

Exceptional mean

time (sec.)

Percentages of outpatients

tested with exceptional time

1 Cataract

2 Glaucoma Visual acuity test 300 40 %

1 Cataract

2 Glaucoma Refraction test 600 10 %

2 Glaucoma Visual field test 500 30 %

TABLE VI. AVERAGE WAITING TIME ESTIMATED FOR EACH DAY OF THE WEEK

Day Ophthalmologist

name

Disease

name

Simulation-based waiting time

(hours : minutes : seconds)

Real waiting time


Monday A Cataract 1:17:46� 1:03:54�

Tuesday B Glaucoma 1:31:25� 1:10:06�

Wednesday A Cataract 1:23:55� 1:08:29�

Thursday A Cataract 1:21:26� 1:14:01�

Friday C Cataract 1:34:07� 1:12:52�

Saturday B Glaucoma 1:30:57 1:06:08�

Sunday E Cataract 1:20:56� 1:13:05�

TABLE VII. CHANGE IN ESTIMATED WAITING TIME WITH THE PASSAGE OF TIME

Outpatients contracting cataract Outpatients contracting glaucoma

Arrival

time Simulation-based

waiting time


Real waiting time


Simulation-based

waiting time


Real waiting time


8:30 0:55:19 0:54:47 0:57:59 0:32:36

9:00 1:21:15 1:02:42 1:19:20 0:51:58

9:30 1:28:49 1:03:34 1:24:08 0:54:50

10:00 1:30:33 1:08:34 1:30:00 1:00:32

10:30 1:42:15 1:15:28 1:29:47 1:00:59

11:00 1:22:18 1:01:54 1:31:26 1:06:54

11:30 1:29:52 1:01:57 1:48:21 1:45:14

13:00 1:35:28 1:14:50 1:20:11 0:56:07

13:30 1:52:31 1:38:02 1:43:37 1:18:58

14:00 1:45:03 1:27:05 1:25:19 1:18:31

14:30 1:46:18 1:18:54 1:34:46 1:08:43

15:00 1:07:13 0:50:32 1:25:33 1:06:29

15:30 1:05:57 0:41:14 1:29:11 1:01:47

16:00 0:57:41 0:45:06 1:12:50 0:48:18

2929

Tracking of Multiple Honey Bees on a Flat Surface

Toshifumi Kimura and Mizue Ohashi School of Human Science and Environment,

University of Hyogo Hyogo, Japan

{kimura, ohashi}@shse.u-hyogo.ac.jp

Karl Crailsheim and Thomas Schmickl Department of Zoology,

Karl-Franzens-Universität Graz Graz, Austria

{ karl.crailsheim, thomas.schmickl}@uni-graz.at

Ryuichi Okada Kagawa School of Phamaceutical Sciences,

Kagawa Bunri University Kagawa, Japan

okd@ kph.bunri-u.ac.jp

Hidetoshi Ikeno School of Human Science and Environment,

University of Hyogo Hyogo, Japan

[email protected]

Abstract— Social insects interact extensively with their mates. To reveal the mechanism of their sociality, it is important to observe the behavior of each individual in a colony and to analyze their social contexts. Recently, digital video cameras have become less expensive despite increases in their performance, enabling researchers to record the behavior in various places easily. However, much labor and time is needed to analyze the social behavior manually from the video footage. In this study, we propose a new method to detect and track multiple bees moving on a flat surface. Our proposed method consists of three core processes: (i) detecting the bee candidate regions using a background subtraction method and binarization, (ii) identifying individuals by a combination of overlapping information in temporal changes of position and simple prediction of the candidate regions based on the bee’s movement, and (iii) outputting the locations and trajectories of the identified bees. Our system succeeds in processing a video of sixteen bees moving freely on a flat arena for three minutes. More than 95% of the bees’ central points were successfully extracted and their trajectories precisely traced.

Keywords-behavioral tracking; ethology, honey bee (Apis mellifera); social insect

I. INTRODUCTION

Social insects, such as honey bees, ants, and termites, interact extensively with their nest mates. In particular, it is common knowledge that honey bees have a developed sociality since the discovery by von Frisch of dances as their communication language [1]. To reveal the mechanism of their social interactions, it is important to record and observe individuals in their social contexts and to analyze their behavior in detail.

Various kinds of recording methods, such as digital video recording, have been developed to measure animal behavior. In recent years, CCD cameras have become less expensive despite increases in their performance. Since video recording is quite effective and an easy way to realize continuous observation, researchers can record videos easily and use these to evaluate the behaviors of the target insects

quantitatively. However, since image analysis of natural scenes depends strongly on the subject and the conditions, these have mostly been conducted by manual operations resulting in much labor and time spent on fairly mundane work. Although automatic tracking methods to analyze multiple individuals have recently been developed by various research groups [2][3][4], it is difficult to apply these methods to analyze the social behavior of an animal group. This is because the methods were originally developed to process the independent behavior of a few individuals, sometimes only in a part, and not the whole, of their movable area in a hive or on another uniform background.

In this study, we present a new method to track multiple honey bees, that are free to walk anywhere on a flat surface. The method is based on our previous method developed to analyze honey bee behavior in an observation hive [5]. Our system can automatically detect the location and movement of the honey bees as numerical data through their identification number (ID) and coordinates.

II. MATERIALS AND METHODS

We used a behavioral video of young honey bees (Apis mellifera) recorded using an infrared video camera in the laboratory at Karl-Franzens-Universität Graz, Austria [6][7]. Sixteen walking individuals in a circular arena were recorded from above. The frame rate and resolution were 25 frames per second (PAL format) and 532×576 pixels, which were relatively low specs, but enough to achieve our aim. The real size of a frame was 600×600 mm with the square of each pixel expressed as 1.13×1.04 mm. The work flow of our proposed method, illustrated in Fig. 1, consists of three core processes: 1) detection of the bee candidate regions using a background subtraction method and binarization, 2) identification of individuals by combining overlapping information in temporal changes of position and simple prediction of the candidate regions based on the bee’s movement, and 3) outputting the locations and trajectories of identified bees at every frame.

In the first process, our system creates a background image from the recorded video. The video file is separated


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.25

36


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.25

36

into sequences of images, where each image consists of three color components, Red (R), Green (G), and Blue (B) with 256 quantized values. These images can be mapped onto a one-dimensional vector space from a 3-dimensional one by a gray-scale transformation without losing any characteristics of the original images, because we used an infrared image in which the R, G, and B values are similar. Next, our system finds the maximum value of each pixel, so that the pixel values of a honey bee are low, while those of the background are high. The background image was created from a one-minute video. The image can be separated into the bees’ region (foreground) and the floor of the arena (background) by subtracting the pixel values of the background image from the original image and using a conventional threshold binarization.

In the second process, individual honey bees are extracted and identified from clustered images. The number of pixels comprising a single bee’s body is determined from the distribution of the number of pixels occurring in individual bee images within 500 frames. Based on the

number of pixels for a single bee’s region, we classify the entire bees’ regions into the following three categories using the temporal information of the overlaps between current and previous bee regions: 1) the region in the current image that overlaps the region of one bee in the previous image (One Bee Region (OBR)), 2) the region in the current image that overlaps the regions of two or more bees in the previous image (Two or more Bees’ Region (TBR)), 3) the region in which no bees occur. Each OBR is assigned a unique number as its ID at the first appearance or reappearance of the image frame. Correspondence of IDs among different frames is determined based on a temporal context of movement. The TBR is divided into several OBRs by considering collision and overlap if the size of its region is greater than a Single Bee Size (SBS). The simple moving prediction is based on the probability that a bee is present in the next frame and we predict that the honey bee moves to the pink areas in Fig. 2. Fig. 2A shows the case where there is no movement of the bee in the previous two time units. The moving area is 5×5 pixels with the bee’s previous position in the center. Figs. 2B and C show cases with right and slanting movements of the bee. In these cases, we predict that the honey bee moves 2 pixels forward, and then 1 pixel back or to the side, respectively. Using these predictions, we calculate the overlapping ratio between the current TBR and each region created by moving the previous OBRs. From the maximum ratio, the current locations of individuals in the TBR are determined and these are numbered with the corresponding ID. These processes are repeated to the end of the frames to track all the movements.

In the final process, the numerical value of each bee’s location in every frame and the trajectories are exported to a CSV (comma separated values) and a video file, respectively. A bee’s location is represented by the central point of its body, while the trajectories are created by temporally connecting points with the same ID.

Figure 1. Work flow of our proposed method. Our system consists ofthree core processing blocks; process (i) is the detection of candidate regionsof bees, process (ii) is the detection and identification of each individual, andprocess (iii) is the visualization of results such as a bee’s location and itstrajectory.

Figure 2. The moving prediction of a bee. (A) The case where a bee doesnot move from a previous to the current time. (B-C) The cases where a beemoves in the same interval, such as right or left turns, skew, and so on. Beesmove in right and upper right directions.

3737

III. RESULTS AND DISCUSSIONS

We developed a prototype system based on the proposed method using Microsoft Visual Studio 2008 (Visual C++) and OpenCV 2.0 [8] on a computer with an Intel Core2Quad 2.83 GHz (CPU), 8 GB (Memory), 1 TB (HDD) and running Microsoft Windows XP Professional 64-bit (OS). Our developed software is executed as a Win32 console application on Windows.

We prepared a three-minute video with 4,500 frames. Our prototype system processed this movie, detected the individuals in each frame, and drew their trajectories. The central points of each frame were output as a CSV file. Fig. 3A shows the 1,500th frame in our experiment. Fig. 3B shows that our system correctly detected the regions containing bees when compared with Fig. 3A. Our system plotted the central points of bees in this frame (Fig. 3C) and connected the points with the same bee’s ID (Fig. 3D) to visualize the tracking results of our system.

We randomly selected three six-second blocks from the original video and manually detected the central points of all bees in these blocks to validate the results. Thereafter, we calculated the detected error as the Euclidean distances

between the points obtained by the automatic method for the three videos and those by the manual method. In most of our measurements, the errors between bee positions detected automatically and manually were less than 2.0 mm (Fig. 4). The average and variance of the total errors were 2.87 mm and 6.24 mm, respectively. If the errors are less than 6.0 mm, we consider that the honey bee position has been successfully extracted, because the bee’s body length is about 13.0 mm. This suggests that our automatic extraction can correctly detect the position of bees in most cases, that is, more than 95% of all central points. We also confirmed that our system is able to track bees in multiple regions with several bees touching, overlapping, and/or crossing. Fig. 5 shows two bees moving from time t1 to t6 with the bees, "a" and "b", represented as red and blue circles, respectively. In this case, our system successfully tracked the bees even though they crossed and overlapped between time t3 and t4. However, there are also several cases where the system failed to track multiple bees. This occurred when walking or rotating individuals overlapped completely with one another (Fig. 6). Our program also failed to track more than three bees with overlapping regions. One of the reasons for these errors is that our system predicts a bee’s movement based on

Figure 3. (A) Original image. (B) Candidate regions of bees detected by our method. (C) Results of identifying individuals. Enlarged image of a part of (C)is shown in the upper right hand corner. (D) Trajectories of all the bees as drawn by our software.

3838

straight line movement without any rotation. To improve our system, the moving prediction of honey bees must also consider turning in the movement.

IV. CONCLUSION

We proposed a new method to track multiple honey bees on a flat surface and applied our method to a video of sixteen bees walking freely on a flat area. More than 95% of the individuals’ locations were extracted and traced.

In the future, we plan to improve our method by adopting a more advanced moving prediction such as an AR model and Hough transform in order to improve the performance of identification and tracking. We also intend to analyze the social behavior of individuals using our developed system.

ACKNOWLEDGMENT

Kimura was partly supported by the Grant-in-Aid for Scientific Research (KAKENHI) (No. 23650157). Schmickl was supported by a FWF research grant “temperature-induced aggregation of young honey bees”, no. P19478-B16. The authors were supported by the following grants: EU-ISTFET project SYMBRION, grant agreement no. 216342; EU-ICT project REPLICATOR, grant agreement no. 216240; FWF project REBODIMENT (project number: P23943-N13).

REFERENCES

[1] K. von Frisch, "The dance language and orientation of bees", Harvard University Press, Cambridge, MA, USA, 1967.

[2] T. Balch, Z. Khan and M. Veloso, "Automatically tracking and analyzing the behavior of live insect colonies", Proc. the Fifth International Conf. on Autonomous agents, pp.521-528, 2001.

[3] A. Feldman and T. Balch, "Automatic Identification of Bee Movement", Proc. 2nd International Workshop on the Mathematics and algorithms of social insects, pp.53-59, 2003.

[4] Z. Khan, T. Balch and F. Dellaert, "Efficient Particle Filter- Based Tracking of Multiple Interacting Targets Using an MRF-based Motion Model", Proc. 2003 IEEE/RSJ International Conf. on Intelligent Robots and Systems 1, pp.254-259, 2003.

[5] T. Kimura, M. Ohashi, R. Okada and H. Ikeno, "A new approach for the simultaneous tracking of multiple honey bees for analysis of hive behavior", Apidologie, 42, pp.607-617, 2011.

[6] S. Kernbach, R. Thenius, O. Kernbach , T. Schmickl, "Re-embodiment of Honeybee Aggregation Behavior in an Artificial Micro-Robotic System, Adaptive Behavior", 17(3), pp.237-256, 2009

[7] T. Schmickl, H. Hamann, "BEECLUST: A Swarm Algorithm Derived from Honeybees. Derivation of the Algorithm Analysis by Mathematical Models and Implementation on a Robot Swarm.", In: Bio-inspired Computing and Communication Networks, Yang Xiao, Dr. Fei Hu (eds.), Auerbach Publications, CRC Press, 2011.

[8] Welcome - OpenCV wiki, Available: http://opencv.willowgarage.com/wiki/

Figure 4. Histogram of Euclidian distance errors between manual andautomatic detection. The red dotted line shows half of a bee’s body length.

Figure 5. An example image of the successful tracking of bees inoverlapping areas. The bees named "a" and "b" are represented as red and blue circles, respectively. The circles for "a" and "b" at time t3 show the locations of the bees when the image was recorded.

Figure 6. The images (A-D) show an example of failed tracking. (A) Twoindividuals are located at separate positions. (B) "a" moves downwards, but"b" remains almost in the same position. (C) "a" drifts from t1 to t2 in astraight line and locates near "b" (red filled circle). On the other hand, ’b’moves to the lower left (blue filled circle). In this situation, the labels of "a"and "b" are switched because of a failure in estimating the movement. (D)Since incorrect labels are continuously assigned to the bees, we obtainincorrect movements.

3939

Retrieval Performance of Complex-Valued Associative Memorywith Complex Network Structure

Kohei Omachi, Teijiro Isokawa, Naotake Kamiura, Nobuyuki MatsuiGraduate School of Engineering,

University of Hyogo,Himeji, Hyogo, Japan

Email: [email protected]

Haruhiko NishimuraGraduate School of Applied Informatics,

University of Hyogo,Kobe, Hyogo, Japan

Email: [email protected]

Abstract—The performances of associative memory, basedon complex-valued Hopfield neural network, are presented inthis paper. The structures and parameters of complex networksare considered in the presented network. Retrieval performancewith respect to the cluster coefficient, one of the parametersin complex networks, are explored.

Keywords-Associative memory, Complex-valued neural net-work, Complex network, Cluster coefficient

I. INTRODUCTION

Complex network is a type of topological networks thatoften appear in the fields of social, biological, communica-tion, and so on [2]. Scale-free networks [1] and small-worldnetworks [5] are typical and well-studied cases of complexnetworks. These networks are known as robust against thedefective elements in the networks and they achieve highefficiencies within their communications. The researches onthese networks have been conducted in order to find theiruniversal properties and also to reveal the phenomena in thereal world from the viewpoint of complex networks.

Though much attention has been attracted on the analysisof the networks in the real world, there seems a smallnumber of researches for utilizing the properties of com-plex networks in designing artificial networks, especially,artificial neural networks [4]. Kim explored the retrieval per-formances on the real-valued Hopfield neural networks, andshowed that these are correlated with the cluster coefficient,that is one of the properties of the complex networks [3].This work is a first step for introducing the structures ofcomplex networks to the artificial neural networks, and fewattempts have been conducted to introduce them to the neuralnetworks with higher dimensions.

This paper presents an associative memory based oncomplex-valued Hopfield neural network. The structures ofcomplex network is then adopted, and the relations betweenthe properties of the networks and the retrieval performancesof the network are explored. This paper is organized asfollows. In section 2, fundamentals of the complex networksand complex-valued associative memory are recapitulated.Section 3 describes the retrieval performances with the

clustering coefficients. Finally, section 4 summarizes ourresults with discussions.

II. PRELIMINARIES

A. Complex network

Figure 1 shows an example of topological networks, con-sisting of so-called nodes and edges. A network is configuredby connecting nodes by edges, as shown in this figure. Inthis network, degree of the node s, denoted by deg(s), isdefined by the number of edges connected to the node s.Also, the distance between the node s and node t, lst, isdefined by the minimum number of edges counted betweennodes s and t.

By using these quantities, the following properties of thenetwork can be introduced. Averaged distance over nodes inthe network with n nodes, denoted by L, is defined by

L =1

n(n− 1)/2·∑

i,j,i�=j

lij . (1)

This means the averaged distance for all the combinations ofpair of nodes. Less values of L imply that the propagationspeed of information from a node becomes higher. Clustercoefficient C is a quantity of the network which representsthe density of connections in the closed subgroup of thenetwork. For calculating the cluster coefficient for the nodes, Cs, firstly the nodes, which are connected to the nodes with only one edge, are selected. The number of thesenodes is the same as deg(s). Cs is then calculated as theratio of the number of pair of nodes that directly connectedover the number of all combination of nodes (ks(ks−1)/2).The cluster coefficient C is then obtained by the average ofCss for all nodes:

C =1

n

n∑

s=1

Cs (2)

Cluster coefficients for complex networks tend to take highervalues, and this implies that the dense connections, or diverseconnections, are likely to exist in each of subgroups in thenetwork. Thus, robustness against the defects on edges isendowed in complex networks.


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.26

40


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.26

40

Figure 1. An example of topological networks with nodes and edges

(a) (b)

Figure 2. Construction of Watts and Strogatz network from a locallymeshed network

(a) original W-S network

(C = 0.53)

(b) After 10 exchanges of

edges (C = 0.4)

(c) After 23 exchanges of

edges (C = 0.2)

(d) After 30 exchanges of

edges (C = 0.0)

Figure 3. Construction of a network with lower cluster coefficient by using edge exchanges

B. Watts and Strogatz model with various cluster coefficients

Watts and Strogatz (W-S) model is one of complex net-work models, proposed in [5]. Networks with small values ofaveraged distance L and large values of cluster coefficientC can be produced according to the simple procedure, asfollows:

1) Prepare n nodes on an one-dimensional ring. Let eachnode connect to k neighboring nodes on its left andright by edges, forming the network (see Fig. 2(a)).

2) Select a pair of edges randomly at the probability p(0 ≤ p ≤ 1). Each of selected edges has its nodes atboth ends, choose randomly either of nodes and detachit.

3) Select randomly a node in the network, except for thenode just detached and the node on the other end ofthe edge, then attach the selected node to the edgewithout a node.

An example of the W-S network with the above procedureis shown in Fig. 2(b). The dotted lines in the networkrepresent the edges undergone the processes of detach andattach. Such edges, called short-cuts, become long-distanceconnection between nodes, and make the averaged distanceof the network L be smaller and thus the propagation ofinformation in the network will be efficient.

Cluster coefficient Cs for networks according to the aboveprocedure show constant values if particular values of n andp are given. For obtaining the networks with various Cs weapply the edge exchanging scheme proposed in [3]. In thisscheme, two pairs of nodes, in which each node of pairs isidentical, are selected and a node on one pair is connected

to a node on the other pair by exchanging the edges ofnodes. The degree of each node still remains constant afterthis operation. For example, if the nodes {a, b, c, d} areprovided with the connections a–b and c–d, the resultantpairs become a–c and b–d. In this paper, the exchange ofpairs is repeatedly conducted under the condition that theresultant Cr becomes lower than the original Co with Co−Cr ≤ 0.01. Figure 3 shows the transition of the network withthe cluster coefficients according to the exchanging scheme.

C. Complex-valued associative memory

A complex-valued Hopfield neural network is used foran associative memory in this paper. Our Hopfield networkconsists of n neurons where the input, output, action poten-tial and connection weights are encoded by complex values.The action potential of the i-th neuron, si is

si =

n∑

j=1

wijxj ,

where xj is the output of the j-th neuron and wij denotesthe connection weight from the j-th neuron to i-th neuron.The neurons are asynchronously updated by

xi = f(si),

where f is an activation function. In this paper, we use f as

f(s) =ηs

η − 1 + |s| ,

where η is a real valued constant satisfying η > 1. The con-nection weights wpq are constructed by so-called Hebbian

4141

rule as:wij =

1

n

∑

µ=1

npεµpε

µp∗,

where εµp represents the pattern vector for the p-th neuronin the μ-th pattern and np is the number of patterns to bestored. Also, εµp

∗ denotes the complex conjugate of εµp . Theconnection weights by the Hebbian rule satisfy wpq = w∗qp.

D. Associative memory with complex network structure

The proposed associative memory is presented in thissection. Its structure is a form of a W-S network and thepatterns are embedded onto the network by using Hebbianrule. The construction scheme is as follows:

1) The network with n nodes with a given cluster coeffi-cient is constructed by the scheme described in SectionII-B.

2) The connection matrix R = {rij} with n×n elementsis prepared for the constructed network, such thatrij = rji = 1 if node i and node j in the networkis connected, otherwise rij = rji = 0.

3) The weight matrix W for the Hopfield network withn neurons is constructed by Hebbian rule with givenstored patterns.

4) The weight matrix W is updated by the connectionmatrix R by wij ← rij · wij .

III. RESULTS

This section describes the performance of our proposedassociative memory through numerical simulations. Theconditions of the simulations are as follows:

• The number of neurons is 15× 15 = 225.• Networks are prepared for the Cluster coefficients C ={0.6, 0.5, 0.4, 0.3, 0.2, 0.1, 0.0}, constructed from W-Snetworks with p = 0.01 (C ≈ 0.68).

• The number of networks with each C is 10.

We adopt two quantities for the performance evaluation. αis the ratio of overlapping of neuron states for the retrievedand the desired patterns. α = 1 is accomplished when theretrieved patterns and the pattern to be retrieved is same.β corresponds to the ratio that the correct patterns can beretrieved for the input pattern.

A. np = 1 case

First we consider the case where only one pattern isembedded to the network, i.e. np = 1. The patterns tobe embedded are composed of four kinds of neuron states,{1,−1, i,−i}, that display by the colors {black, white, darkgray, light gray}. Figure 4 show three patterns that are usedfor embedding to the network. A four-tuple number with therounded brackets for each pattern represents percentages offour states in the pattern. These percentages for the pattern#1 are biased rather than the ones for the patterns #2 and#3. In the patterns #2 and #3, the percentages are similar

(a) Pattern #1,(0.14, 0.73, 0.04, 0.10)

(b) Pattern #2,(0.27, 0.27, 0.27, 0.20)

(c) Pattern #3,(0.24, 0.27, 0.22, 0.27)

Figure 4. Three stored patterns with four kinds of states used in thenp = 1 case. Four-tuple number represents percentages of four states inthe pattern

Figure 5. C-dependence of overlap rate for each neuron in the np = 1case

Figure 6. C-dependence of correctness of retrieved patterns in the np = 1case

but the arrangement of the states is random in the pattern#3. Input patterns to the network are created from one ofthe stored patterns with randomly selected 30% of the statesbeing changed by multiplying −1.

Figure 5 shows the transition of α for the cluster coef-ficient for three kinds of the stored patterns. Each α is anaveraged value for 10 different networks with 10 differentinitial states of the network. Similar tendencies are shownfor these stored patterns, and the complete retrievals areaccomplished in the ranges of C ≤ 0.2. This result iscorrelated to the result obtained by the real-valued Hopfieldnetworks [3].

The transition of β is shown in Fig. 6. This shows similartendencies to that of α, but the percentages at higher Csare comparatively low. This is because β is calculated as

4242

(a) Pattern #1

(b) Pattern #2

(c) Pattern #3

(d) Pattern #4

Figure 7. The stored patterns with four kinds of states used in the np > 1cases

the ratio for the cases of α = 1.0, thus a small number ofneuron states are settled in the wrong pattern.

Both values of α and β show that networks are not goodat storing the pattern #3. The neuronal states in the pattern#3 (see Fig. 4(c)) are randomly placed, so there are norelevance between neuron states and their locations. Thenetworks are configured as locally connected by producingthe W-S network structure, thus it is difficult to embed therandomly distributed patterns.

B. np ≥ 1 cases

We proceed to embed the multiple (np ≥ 1) patterns ontothe networks. The stored patterns are shown in Fig. 7. Thesepatterns are orthogonal each other, i.e., the correlationsbetween any two patterns are set to 0. This is due to thelimitation of the Hebbian rule to embed the pattern.

We investigate the retrieval performances in the cases ofnp = 1 (pattern #1 is embedded), np = 2 (patterns #1 and #2are embedded), and np = 4 (all the patterns are embedded).The transition of α for the cluster coefficient C is shown inFig. 8. From this result, unstability for embedding patternsappears for np = 2 and np = 4 cases even if C takes lowervalues. The changes of β are also shown in Fig. 9, and thesealso show similar tendencies to the changes of α.

IV. DISCUSSION AND CONCLUSION

This paper presents an associative memory with thestructure of complex networks, based on complex-valuedHopfield neural network. The patterns are stored by theHebbian rule and the connection of neurons are designedby the modified Watts and Strogatz network model. The as-sociative performances are investigated with several numberof patterns being stored.

From the numerical simulations, it is found that theretrieval of patterns is successful with low values of thecluster coefficient. From the viewpoint of computationalcost for updating the network, due to many connectionweights being lost during the process of constructing thenetwork, thus the low computational cost is achieved withmaintaining its retrieval ability. During decreasing the clustercoefficient of the network, the edges (connection weights)are re-connected over the network, so the efficient retrievalscan be accomplished.

As for our future work, the other types of properties ofcomplex network, such as L, are adopted for evaluating the

Figure 8. C-dependence of overlap rate for each neuron in the np ≥ 1cases

Figure 9. C-dependence of correctness of retrieved patterns in the np ≥ 1cases

retrieval performance. Evaluations on larger scaled networks(n > 1000) and comparisons with randomly spanned net-works will be necessary.

ACKNOWLEDGMENT

This study was financially supported by Japan Society forthe Promotion of Science (Grant-in-Aids for Scientific Re-search (C) 23500286) and Young Scientists (B) 24700227).

REFERENCES

[1] A. Barabasi and E. Bonabeau, “Scale-free networks,” ScientificAmerican, vol. 228, pp. 50–59, 2003.

[2] A. Barrat, M. Barthelemy, and A. Vespignani, Dynamicalprocesses on complex networks. Cambridge University Press,2008.

[3] B. J. Kim, “Performance of networks of artificial neurons:The role of clustering,” Physical Review E, vol. 69, no. 4,pp. 045 101–1–045 101–4, 2004.

[4] K. Tsuruta, Z. Yang, Y. Nishio, and A. Ushida, “Small-worldcellular neural networks for image processing applications,” inProceedings of European Conference on Circuit theory andDesign, vol. I, 2003, pp. 225–228.

[5] D. J. Watts and S. H. Strogatz, “Collective dynamics of ‘small-world’ networks,” Nature, vol. 393, pp. 440–442, 1998.

4343

Abstract—This paper describes a development of MEMS

(Micro Electromechanical Systems) -based electrostatic energy

harvester for scavenging energy from ambient vibration.

Generally, power output of electret harvester is increased with

the narrower air-gap between an electret and an opposite

electrode. However, electrostatic attraction force between the

electret and the opposite electrode is also increased. Hence the

gap control is crucial to avoid the pull-in. We measured the

displacement of the vibrating mass moved by electrostatic force.

As a result, the electric discharge between the Si grid and the

base electrode (BE) is occurred when the BE was supplied over

110 V. The vibrating mass is pulled-in at 150 V if the discharge

does not occur at the potential. Therefore we proposed and

design the electrostatic energy harvester using the micro spacer

to prevent the pull-in situation. The size of the spacer is 86 × 86 ×

27 µm3 to prevent the mass fixed by the friction.

Index Terms—Electret, Energy Harvesting, Pull-in, Spacer

I. INTRODUCTION ECENTLY, autonomous sensing systems have attracted considerable attention in various applications [1]. These systems required a small size and a long life power source

for those continuous operations. Thus, an energy harvesting power source is the most important and critical device for human monitoring. Electrostatic power harvesting mechanism based on electret is very suitable for low frequency vibration such as human movement [2]. Electret harvesters were fabricated by combining an electret and an opposite electrode after the charging step [3]. In order to improve the yield of the fabrication process and the performance of the harvesting, we developed the selective charging method by using the buried grid electrode (BGE) [4] and Si grid electrode [5] shown in Fig. 1. The Si grid with fine slits acts as the grid electrodes in the charging process, and as the opposite electrodes in the harvesting operation [6]. We aim at the realization of the energy harvesting from the human motions. Hence the moving structure requires soft springs to achieve a long-range displacement at a low frequency. Fig. 2 shows the image of the energy harvester. The generator consists of three layers; the electret part with

K. Fujii is with the Dept. of EECS Graduate School of Engineering, University of Hyogo (e-mail: [email protected]).

movable structure, the Si grid and the cover glass. The chip size is 12 × 13 × 0.5 mm3 with the movable mass (10 × 10 × 0.5 mm3) and the single bended double springs with width of 20 µm. The aspect ratio of the beam is 20. The movement of the structure for the X, Y, and Z direction with the applied acceleration of 1 G is 48 μm, 0.5 μm, and 0.4 μm, respectively. The simulated resonant frequency for the X direction is calculated to be 71 Hz by using finite element method. In this study, we employ the selective charging method by using the BGE and the Si grid electrode, and measured an action of the movable mass that moved by electrostatic force. The micro spacer for preventing the pull-in is also presented, which is fabricated by MEMS (Micro Electromechanical Systems) batch process.

Fig. 1. Schematic of the electret based generator using Si grid as the grid electrode in the charging process.

Fig. 2. Schematic of the electret based generator.

Development of Electret Energy Harvester with Micro Spacer

Kohei Fujii, Toshikazu Onishi, Non-Member, IEEE

Koji Sonoda, Takayuki Fujita, Kensuke Kanda, Kohei Higuchi, and Kazusuke Maenaka, Member, IEEE

R

Slit

Mass

Pad Si grid

Electret & Electrode

Pad space

Spring Vibration

Electret part

Cover grass

Cytop

Cytop

Mass

glass


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.16

48


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.16

48

II. CHARACTERISTICS OF ENERGY HARVESTER Fig. 3 shows the photograph of the electret energy harvester bonded on a PCB (Printed Circuit Bonded). The chip size is 12 × 13 mm2 with the movable mass (10 × 10 × 0.5 mm3) that is supported by the single bended double springs. The beam width decreases from 20 µm to 12 µm for the bottom side and 16 µm for the top side of the trapezoid cross section because of the over etching in the DRIE (Deep Reactive Ion Etching) process. Fig. 4 shows the measurement setup for the load characteristic. The energy harvester is fixed on the shaker, after the electret was charged with -30V. The excitation amplitude and frequency are 200 µm peak-to-peak and 20 Hz, respectively. The output is buffered to eliminate parasitic elements by ultra high-input impedance, Zin OpAmp after a variable load resistance. The output power was calculated from the measured voltage using digital multimeter (DMM) and the load resistance. Fig. 5 shows the optimized output power by the impedance matching. The output power was 1.8 nW with the optimal load resistance of 32 MΩ.

III. MEASUREMENT Since the power output is proportional to the square of the surface charge density of the electret, the high potential of it is required. However, the electret and the opposite electrode are pulled each other by an electrostatic force. The pull-in makes the output power to be unstable because the electret and the opposite electrode are fixed. We calculated and measured the displacement of the oscillation mass moved by electrostatic force. We estimate the relation of supplied voltage to BE (base electrode, beneath the electret film) and displacement of the mass in Z direction by utilizing the equation of electrostatic force and repulsive force of the spring:

(1) where A is the overlap area of the electrode, ε0 is the vacuum permittivity, d0 is the initial air-gap, Δd is the displacement of the moveable mass, and K is the spring constant. The harvester was fixed on the vibration-proof stand. The electrical potentials on the Si grid, the BE and the BGE were 0 V, 0 to 110V and 0 V, respectively. The moveable mass came close to the Si grid due to the electrostatic force, which attracts the Si grid and the BE when the supplied voltage is increased. Fig. 6 shows the measurement results of the displacement of the oscillation mass in Z direction versus the supplied voltage to the BE that measured by using the laser interferometer (New view 6300, Zygo Corp, CT, USA). The theoretical values were calculated from (1). Table 1 shows the dimensions and spring constants of the harvester. The Z-directional spring constant, KZ was calculated

from the measurement results, and the X-directional spring constant KX was calculated from resonant frequency and the mass M.

Fig. 3. Photograph of the electret energy harvester bonded on the PCB.

AD549 Ultra high-Z OPAZin > 1013Ω 100 Ω

Energy Harvester 10 MΩ - 1 GΩLoad resistance

Shaker withLaser range finder

DMMZin = 1 MΩ

Fig. 4. Measurement setup for load characteristics.

Load resistance [Ω]

Out

put p

ower

[W]

10-10

10-9

10-8

106 107 108 109

32 MΩ

1.8 nW

Roptimal

Poptimal

Fig. 5. Output power versus load resistance. Excitation frequency and vibration amplitude are 20 Hz and 200 µm, respectively.

V =2K (Δd )

ε 0A(d0 − Δd)

13 µm 12 µm

BGE

BE

Si grid

BG

Si Si

BEDevice

PCB

4949

Fig. 6. Measurement of the displacement of the mass.

IV. DISCUSSION AND FUTURE WORK

A. Effect of electrostatic force

The electric discharge between the Si grid and the BE occurred when the applied voltage to the BE over 110 V. The discharge is caused in the unprotected electrode on the spring. In order to prevent the discharge between the electrodes, the CYTOP (CTL-809M; Asahi glass Co., Ltd. Japan) can be spun on the electrode. The oscillation mass is pulled-in at 150 V if the discharge does not occur at 110 V. The pull-in voltage was different from the theoretical curve because of a non-ideal spring structure. Moreover, the overlapping area between the Si grid and the BE was attracted by the applied voltage because the moveable mass was attracted toward X-axis by the electrostatic force. As a result, the distance of the air-gap decreased and then pull-in occurred.

B. Pull-in prevention structure

The electret should be charged with the higher potential to obtain more output power. However, the electret and the opposite electrode are pulled with each other by electrostatic force. A spacer is adopted to prevent a sticking between the electret and the electrode. Fig. 7 shows the layout of the micro spacers. The size of the spacer is 86 × 86 × 27 µm3, which is as wide as the slit width of the Si grid. Theoretically, The pull-in is occurred when the air gap decrees to the two-thirds of the initial gap. Therefore the height of the spacer is determined at 27 µm. The area of the spacer is narrow to reduce the area of contact with the mass. The spacers are located at four points near the corners of the mass. The SiO2 on the spacer prevents the discharging between the spacer and the BE.

Fig. 7. Layout of micro spacer.

D. Process flow

Fig. 8 shows the fabrication process of the harvester. For the electret part, the 86 µm-width BE and the 94 µm-width BGE were patterned on a 525 µm-thickness Si substrate with 1 µm-thickness SiO2 film. The width of the BE and the BGE were optimized to the maximum change of the capacitance for the 100 µm peak-to-peak mass traveling [4]. The 3 µm-thick CYTOP film was spun on the Si substrate and patterned by O2 plasma etching. Then, the movable structure was formed by the DRIE. For the Si grid, SiO2 mask on a 300 μm-thick p-type Si substrate (< 0.1 Ω-cm) was patterned for the 2nd

Element Quantity

Spring width 12 µm for top 16 µm for bottom

Spring length 3.8 mm Spring

constant

X-axis:KX

4.36 N/m

Spring

constant

Z-axis:KZ 906 N/m

Element Quantity

Mass:M 1.1510-4 kg Electrode

Overlap

area:A 4.3×10-5 m2

Electrode

length 10 mm

Electrode

widgth 40 µm

Initial gap:d 40 µm

27 µm

Spacer

Electret part

Si grid part

UNITS FOR DEVICE PROPERTIES TABLE I

Measured displacement

Theoletical displacement

5050

DRIE process. The Al electrode was sputtered on the wafer and performed sintering process at 460ºC in order to improve the electrical connecting between the Al electrode and the Si grid electrode. After that, The Al electrode was patterned. The 1st DRIE process fabricated the slits of Si grid. The 13 μm-thickness KMPR film was spun on and patterned. The 2nd DRIE process fabricates the 27 µm-thickness spacer by using SiO2 mask. The 2nd DRIE process also fabricates the air-gap. In the assembly process, each part was stacked and bonded by the CYTOP heat-pressing method at 140ºC. Finally, the assembled device was charged by corona discharging with –8 kV for needle electrode and –30 V for the Si grid and the BGE. The BGE set to the same voltage of the grid electrode prevent the electric field between the BGE and the grid electrode. Then the CYTOP will be charged with fine slit pattern as an electret [7].

V. CONCLUSION In the present study, the supplied voltage to the BE

versus the displacement of the oscillation mass in Z direction is measured. The electrical discharge occurred between Si grid and BE at 110 V. The actual pull-in voltage is lower than theoretical pull-in voltage. In order to prevent the discharge between the electrodes, the CYTOP can be spun on the unprotected electrode. We proposed the electrostatic energy harvester using the micro spacer structure to prevent the pull-in. The size of the spacer is 86 × 86 × 27 µm3. The SiO2on the spacer prevent the discharge at the spacer and the BE.

ACKNOWLEDGEMENT This research is supported by the Exploratory Research

for Advanced Technology (ERATO) scheme produced by Japan Science and Technology Agency.

REFERENCES [1] Samson D, Kluge M, Becker T, Schmid U 2010 Energy Harvesting for

Autonomous Wireless Sensor Nodes in Aircraft Procedia Engineering

for Eurosensors XXIV (Linz, Austria, 5–8 September 2010) 4 pages on CD-ROM

[2] P. D. Mitcheson et al., “Energy Harvesting From Human and Machine Motion for Wireless Electronic Devices,” Proceeding of the IEEE, Vol. 96, No. 9, September 2008, pp. 1457–1486.

[3] Y. Naruse et al., “Electrostatic Micro Power Generator From Low Frequency Vibration Such As Human Motion,” Proceeding of PowerMEMS 2008 + μEMS 2008, Sendai, Japan, November 9-12, 2008, pp. 19–22.D. Miki et al., ”MEMS ELCTRET GENERATOR WITH ELECROSTATIC LEVITATION,” Proc. of PowerMEMS 2009, pp.169-172

[4] T. Toyonaga et al., ”Novel Corona Charging Method for Electret Power-Harvester Using Buried Grid-Electrodes,” Tech. Dig. of the 5th APCOT 2010, p.290

[5] T. Fujita et al., ”Selective Electret Charging Method for Energy Harvesters Using Biased Electrode,” Proc. of Eurosensors 2010, B4P-F-3, 4 pages on the CD-ROM

[6] K. Fujii et al., “Electret Based Energy Harvester Using A Shared Grid Electrode” Proc. of Transducers 2011, 2634 pages on the CD-ROM

[7] T. Fujita et al., ”Electret Based Energy Harvester by Using Silicon Grid Electrode,” Proc. of PowerMEMS 2010, pp.407-410

(a) Electret part

(b) Si grid part (c) Assembly

Fig. 8. Fabrication process.

5151

A b s t r a c t — T h i s p a p e r p re s e n t s a w e a r a b l e P P G (photoplethysmographic) sensor system with a PSoC microcontroller unit (Cypress Semiconductor Corp., USA). We have developed a pulse wave monitoring system on earlobe, which has a PIC microcontroller and analog circuits. However, analog sensors need some discrete parts, resulting in difficulty in reducing the system size. The PSoC microcontroller includes a CPU, analog and digital blocks to configure mixed-signal circuits, and an internal oscillator in a single IC package. Therefore, PSoC allows designers to reduce the number of external discrete devices and the system size. The novel system consists of a photo interrupter, the PSoC, and the Bluetooth module, which are on PCBs (Printed Circuit Board) of 18, 13, and 10 mm in diameter, respectively. The photo interrupter detects the pulse wave due to the change in volume of blood vessels. The PSoC includes an analog filter, amplifiers, and digital circuits to calculate HR (heart rate) from the pulse wave of human body. The system was able to detect the pulse wave and transmit the HR data to a PC wirelessly. The power supply voltage is 3.3 V, and the total current consumption is about 51 mA. In comparison with the previous work, the area of discrete RCL (resistor, capacitor, and inductor) devices is reduced from 1230 mm2 to 466 mm2, and the number of such devices is reduced from 26 to 6.

Index Terms—Photoplethysmograph (PPG), Heart Rate (HR), Programmable System-on-Chip (PSoC), Wireless Sensor Network (WSN).

I. INTRODUCTION

Health monitoring systems are useful in the medical field to

analyze human health conditions and it is also hoped that they will become available for personal use as well. WSNs

(Wireless Sensor Networks) are useful to remotely monitor a

person who lives in a distant location [1-6]. Hence, monitoring systems and WSNs are frequently integrated together. A

typical monitoring system is composed of various sensors such as ECG (Electrocardiogram), blood pressure sensor,

accelerometer for movement activity, and the pulse wave sensor. The pulse wave sensor can be useful for indicating the

presence of some diseases. For instance, cardiac diseases can be diagnosed by monitoring the pulse wave. The HR can be

calculated by measuring the cycle of the pulse wave in arterial blood pressure using optical, pressure, ultrasonic Doppler and

laser Doppler monitoring systems [7][8]. In order to use the

system easily in everyday life, a small size and maintenance-free operation are required for the system. Some types of

health monitoring systems such as ring shaped and earrings have been reported [9][10]. These systems have some sensors

and general ly have a microcontrol ler (e .g . PIC

microcontroller; Microchip Technology, USA) in order to control operations and data processing. However, analog

sensors need some discrete components, resulting in difficulty in reducing the system size.

Previously, the authors have developed a pulse wave

monitoring system [11]. The system had a PIC microcontroller and analog circuits that resulted in too large of a size (35 × 35

mm2) for use in everyday life, although it was able to be attached on the earlobe.

The PSoC microcontroller includes a CPU, analog and digital blocks to configure mixed-signal circuits, and an

internal oscillator in a single IC package. Analog and digital blocks of the PSoC can form various modules such as SPI

(Serial Peripheral Interface), UART (Universal Asynchronous

Receiver/Transmitter). Therefore, PSoC allows designers to reduce the number of external discrete devices and therefore

the system size. The goal of this study is the fabrication of small-sized

system, which is to be attached on the earlobe. Therefore, it is important that the size is miniaturized in order for the system

to be worn on earlobe. The authors reported an integrated photo interrupter [12] and a hybrid wafer for a multi-layer

integration technique [13]. These techniques can reduce the device size.

In this paper, we present a wearable PPG sensor system

with a PSoC microcontroller for HR monitoring. We determined that the sensor system could successfully detect

the pulse wave and calculate HR.

II. SYSTEM SPECIFICATION

A. Previous work

The authors have designed a pulse wave monitoring system as shown in Fig. 1. This system has a photo interrupter, RF

module, PIC microcontroller, battery holder, and clip. The PCB (Printed Circuit Board) size is 35 × 35 mm2. The system

was attached on the earlobe and was able to transmit the HR

data to PC. However, the system was too large to wear on the earlobe during everyday life. Many discrete devices needed

for an analog filter, amplifier, and ICs were in the system, increasing the size and weight of the device. In order to reduce

the system size, reducing the number of discrete devices was required. The analog components were not able to be easily

changed because the analog circuits were off-the-shelf devices. Therefore, PSoC microcontroller was selected to

solve this problem.

Koji Sonoda, Yuji Kishida, Tomoya Tanaka, Kensuke Kanda, Takayuki Fujita, and Kazusuke Maenaka

Dept. of EECSGraduate school of Engineering, University of Hyogo

2167 shosha, Himeji, Hyogo 671-2280, Japan

Wearable Photoplethysmographic Sensor System

with PSoC Microcontroller

Kohei Higuchi

Maenaka Human-Sensing Fusion Project, ERATOJapan Science and Technology Agency

2167 shosha, Himeji, Hyogo 671-2280, Japan


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.19

61


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.19

61

Fig. 1. Our previous work. System has a photo interrupter, RF module, pic microcontroller, battery holder, and clip on PCB (35 × 35 mm2). Although system was attached on earlobe, system was too large to monitor continuously

in everyday life.

B. Hardware Description of Novel System

The novel system consists of a photo interrupter (SG-105;

Kodenshi Corp., Japan), a PSoC (CY8C29466; Cypress Semiconductor Corporation, USA), and a Bluetooth module

(KC-22; KC Wirefree Inc., USA), which are on PCBs (Printed Circuit Board) of 18, 13, and 10 mm in diameter, respectively.

Figure 2 shows the photo image of the developed PSoC wearable system. The PSoC includes an analog filter,

amplifiers, and digital circuit to calculate a HR (Heart Rate) from the pulse wave of human body. The PSoC controls the

photo interrupter and the Bluetooth module. The system is designed to be attached on earlobe as shown in Fig. 3.

Fig. 2. Three PCBs are connected with wire. System consists of photo interrupter, a PSoC, and a Bluetooth module, which are 18, 13, and 10 mm in diameter, respectively.

Fig. 3. System is attached on earlobe of the subject by magnets.

C. Circuit

Figure 4 shows the circuit diagram of the system. The power supply voltage is 3.3 V. The photo interrupter detects

the pulse wave due to changes in the volume of the blood vessels. The PSoC amplifies the obtained pulse wave from the

photo interrupter through a LPF (Low-Pass Filter) and

converts the analog signal to digital by using an ADC (Analog to Digital Converter). The HR is then calculated with the

PSoC and the HR data are sent to the Bluetooth module. The Bluetooth module sends the data to PC continuously.

Fig. 4. Photo interrupter (SG-105), PSoC (CY8C29466) and Bluetooth module (KC-22) are used in the system. Supply voltage is 3.3 V.

6262

D. PSoC Block Diagram

Figure 5 shows the PSoC block diagram. The pulse wave is passed through a PGA (Programmable Gain Amplifier) to

amplify the small signal. The comparator outputs a high voltage (supply voltage) when the input voltage is more than

Vsupply/2. The Timer32 module captures the pulse signal from

the comparator. The HR is calculated by capturing the interval between the pulses and sent to the transmitter module TX8 and

the Bluetooth module transmits the data at 9600 baud. The Bluetooth module only transmits data to PC and does not

receive data from the PC.

Fig. 5. Block diagram of PSoC. PGA module amplifies the pulse wave. Timer32 module captures the comparator signal. HR is sent to transmitter

module TX8 and the Bluetooth module transmits the data at 9600 baud.

III. EXPERIMENT AND RESULT

A. Measurement of Heart Rate

In the PSoC, the HR is calculated by measuring the interval

between the pulse waves. Figure 6 shows the pulse wave of

the subject’s finger (male, 22 years old) through the two PGA modules. The comparator outputs high voltage (supply

voltage) when the voltage of the pulse wave in Fig. 6 is more than Vsupply/2 (=1.65 V). The HR data are received by the PC

through Bluetooth and graphed as shown in Fig. 7. The “observed value” data were obtained by manually calculating

the interval of the waveform from the oscilloscope and the “output value” data were outputted on the PC. Figure 7

indicates a HR of approximately 80-100 bpm during measurement but some irregular data were observed.

Fig. 6. Pulse wave of subject’s finger (male, 22 years old) through PGAs. Comparator outputs the high voltage (supply voltage) when the voltage of the

pulse wave in Fig. 6 is more than Vsupply/2.

Fig. 7. HR data show approximately 80-100 bpm during measurement but some irregular data were observed.

B. Current Consumption

The current consumption of the entire system is about 51 mA. The Bluetooth module consumes 25 mA (49%), almost

half of the entire system. The PSoC consumes 19 mA (37%). The photo interrupter consumes 7 mA (14%). The system can

be driven for 1.5 hours with a PD2032 lithium ion cell (75 mAh).

C. Miniaturization with Bare PSoC Die

Figure 8 shows an image of the comparison between the

packaged PSoC IC and the bare (unpackaged) PSoC die on PCB. The system size can be reduced if the packaged PSoC is

replaced with bare die as shown in Fig. 8. The bare die size is

4.6 × 3.3 mm2 which is 28% of the packaged PSoC. The size of the PCB for the bare die is 10 mm diameter. The bare die

can be wired by wire bonding and potted with epoxy resin.

6363

Fig. 8. Packaged PSoC and bare die on PCB. Bare die is 4.6 × 3.3 mm2 and

28% the size of the packaged PSoC. Size of PCB for bare die is 10 mm diameter.

D. Comparison Between Previous Work and Current Work

Table I shows the comparison between the previous work

and the current work. The number of on-chip RCLs was reduced from 26 to 6. The RF module was replaced with

general-purpose Bluetooth and the communication distance was extended from 3 to 10 m, although the current

consumption was increased. The area of the system was reduced from 1230 mm2 to 466 mm2. The area reduction rate

was 62%.

TALBE I. Comparison Between Previous and Current Work

Previous work Current work

CPU pic16f886 PSoC

Chip RCL 26 6

RF module 315 MHz Bluetooth (class 2)

Communication distance 3 m 10 m

Size 35 × 35 mm2 18, 13, 10 mm dia.

Area 1230 mm2 466 mm2

Current consumption 22 mA 51 mA

IV. DISCUSSION AND FUTURE WORK

The system was able to detect the pulse wave and transmit the data to the PC by wireless. The size of the system is small

for attaching the system on earlobe because of the discrete devices reduction. However, the system has two problems that

one is an irregular data observation and the other is high current consumption. The irregular data are sometimes

observed, especially when the subject is moving. The problem may have been caused by each subject, place of body, or

conditions. In the future, some types of attachment will be

investigated such as a magnet and adhesive tapes.The current consumption of the system is 51 mA. The

module is useful for communication with many devices (e.g. PCs, smartphones, and car navigation systems), although the

Bluetooth module occupies almost half of the current consumption of the entire system. The current consumption of

the Bluetooth module can be reduced if the system only intermittently operates the Bluetooth module. The bottom

surface of the PCB for Bluetooth module is empty, so that the

PSoC bare die and the photo interrupter can be placed there. Figure 9 shows a final image of the system. All devices can be

included on a single PCB, reducing the area to 314 mm2.

Fig. 9. Final image of the system. Bottom surface of PCB for Bluetooth

module is empty, so that all devices can be included on a single PCB.

V. CONCLUSION

A wearable photoplethysmographic sensor system with PSoC microcontroller is presented. The PSoC is utilized to miniaturize the system and reduce the number of discrete components. In comparison with the previous work, the discrete RCL devices were reduced from 26 to 6. The RF module was replaced with a general-purpose Bluetooth and the communication distance is extended from 3 to 10 m. The system consists of three PCBs with diameters of 18, 13 and 10 mm. Furthermore, the bare PSoC die was utilized to reduce the PCB size further (13 mm to 10mm diameter). The area of the system was reduced to 466 mm2. The HR data are received on the PC through the Bluetooth interface. The system was able to detect the pulse wave and transmit the data to the PC. The current consumption of the entire system is about 51 mA. The system can be driven for 1.5 hours with PD2032 lithium ion cell (75 mAh).

REFERENCES

[1] A. Darwish and A. E. Hassanien, “Wearable and Implantable Wireless Sensor Network Solutions for Healthcare Monitoring” Sensors 2011, 11, 2011, 5561-5595.

[2] J. Allen, “Photoplethysmography and its application in clinical physiological measurement” Physiological Measurement 28, 2007, R1-R39.

[3] H. Nakamura and T. Tsuji, “Development of Sports Monitoring System” Micromechatronics 48(3), 2004, pp.34-41. (in Japanese)

[4] C. H. Ko, et al., “Multi-sensor wireless physiological monitor module” Proceedings 56th Electronic Components and Technology Conference, 2006, pp.673-676

[5] K. Asakawa, et al., “The construction of physiological information transmission systems by the sensor network” IEICE technical report. Office Information Systems 108(462), 2009, pp.49-54. (in Japanese)

[6] Y.-H. Lin, et al., “A Driver’s Physiological Monitoring System Based on a Wearable PPG Sensor and a Smartphone” Communications in Computer and Information Science 223, 2011, pp.326-335.

[7] T. Kimura, et al., “Wearable Pulse Wave Sensor” IEICE technical report. ME and bio cybernetics 100(350), 2000, pp.43-49. (in Japanese)

[8] U. Rubins, et al., “Real-Time Photoplethysmography Imaging System” 15th NBC on Biomedical Engineering & Medical Physics, IFMBE Proc. 34, 2011, pp.183-186.

[9] B.-H. Yang and S. Rhee, “Development of the ring sensor for healthcare automation” Robotics and Autonomous Systems 30, 2000, pp.273-281.

[10] J. Spigulis, “Optical noninvasive monitoring of skin blood pulsations” Appl. Opt. 44(10), 2005, pp.1850-1857.

6464

[11] M. Ishii, et al., “3D Integrated Pulse Wave Sensor for Miniaturized Pulse Monitoring System” Proc. of the Physical Sensors, IEEJ, 2011, PHS-11-020. (in Japanese)

[12] M. Ishii, et al., “A Pulse Wave Sensor with Ring Shaped photodiode for Miniature Pulse Monitoring System” Proc. of the Japan Ergonomics Society, Kansai Branch, 2011, pp.15-16. (in Japanese)

[13] K. Sonoda, et al, “Hybrid Wafer for Multi-layer Integration MEMS Devices” Proc. of the World Automation Congress (WAC/IFMIP 2010), 2010, No. 461.

6565

Wafer level packaging for MEMS Geiger counter

Asei Miyamotoa, Taichi Hashimotoa,b, Kenichi Makimuraa,b, Kensuke Kandaa,b, Takayuki Fujitaa,b and Kazusuke Maenakaa,b

a) Graduate School of Engineering, University of Hyogo

b)Japan Science and Technology Agency, JST ERATO Maenaka Human Sensing Fusion Project

� Abstract—The size of conventional Geiger counters is large

because the detector (GM tubes: Geiger muller tubes) is fabricated by mechanical processing. In this paper, an MEMS (micro electro mechanical systems) GM tube using a wafer level packaging is proposed. The MEMS-GM tube with glass/Si/glass stacked structure was fabricated by using anodic bonding and polymer bonding. With the supplied voltage of 2.6 kV to the MEMS-GM tube, pulse waves were observed when a radioactive material approached to the fabricated device.

Index Terms—MEMS, Geiger counter, wafer level packaging

I. INTRODUCTION HE radioactive material has been spread widely by the occurrence of nuclear accident due to the Great East Japan

Earthquake. Because human beings have no sensitivity to the radioactivity, it is important to avoid unconscious exposure to the radiation. Interest in the radioactive materials contained in foods has enhanced the demand for measuring instruments of the radioactivity. General radiation detectors on the market are classified based on the sensing principle into Geiger counter with Geiger Muller (GM) tube, semiconductor detectors using ionization and scintillation type detectors using fluorescence [1]. Although only the Geiger counters can detect alpha ray, GM tube incorporated into the Geiger counter is large and expensive, because GM tubes are made by mechanical fabrication process. To achieve the minimization and reducing the fabrication cost of the GM tubes, novel MEMS-GM tube with glass/Si/glass stacked structure realized by wafer-level packaging technology is proposed. A prototype MEMS-GM tube made of silicon and glass wafers, and the evaluations are reported in this research.

II. PRINCIPLES OF GEIGER COUNTERS Figure 1 shows the principle of a conventional GM tube. The

GM tube has a cylindrical metal tube (cathode) and a central electrode (anode). An inactive gas medium is confined into the metal tube and sealed with low pressure using a thin glass window. A voltage of several kV is applied between the central electrode and the metal tube. When a radioactive ray strikes the inactive gas molecules, the gas molecules are ionized and the attracted to the anode. The radioactive ray can be detected by measuring the current generated by the ionized gas molecules. The radiation dose per time can be obtained by counting the current pulse [2,3].

For the MEMS-GM tube, similar principle is employed as illustrated in Fig. 2. The silicon cavity and central pillar are formed as alternatives to the metal cylinder and central electrode (i.e. outside of the silicon cavity is the cathode and central pillar is the anode). As mentioned above, for the MEMS-GM tube, the radioactive ray is detectable by measuring the current derived from the ionized gas molecules. The sensitivity of the GM tube depends on the number of ionized gas molecules and amount of incident radioactive rays inside the GM tubes. Therefore, the cavity should be large size and the window should have low attenuation property with the radioactive ray. The size of MEMS-GM tube chip is determined to be 15 mm × 15 mm for its portability.

Asei Miyamoto is with Graduate school of Engineering, University of

Hyogo, Japan (corresponding author, e-mail: [email protected] ).

T

Inactive gas RadioactivityCathode

RadioactivityAnode

Window

Fig.1 Structure of conventional Geiger counter.

Anode Cathode

Silicon Glass 15 mm

500-1000 �m Glass (window)

15 mm Fig.2 Structure of MEMS-Geiger counter.


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.20

66


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.20

66

III. WINDOW MATERIAL FOR MEMS-GM TUBE Traditional GM tubes use simple monatomic or diatomic

gases, or mixtures of simple gases with small percentages quenching admixtures as the detecting gas of the radioactive ray. To obtain the high sensitivity, the gases with the pressure range in 10 to 27 kPa are encapsulated into the cavity, and a mica or thin glass is used as the window [4]. A radiation dose is attenuated by window materials because a radioactive ray is particles or high frequency electromagnetic wave. Typically alpha and beta decays occur in radioactive materials. The alpha decay radiates an alpha particle (ray) and a gamma ray. The beta decay radiates beta ray and gamma ray. Because the size of alpha particles (ray) is large, the alpha ray is immediately blocked out by the window materials. Beta particles (ray) of electron or proton are slowly attenuated. The gamma ray of electromagnetic wave is attenuated more slowly than the beta ray by window materials. Since the window material should have properties with low attenuation for the radiation dose, a borosilicate glass (cover glass, Company MUTO PURE CHEM) with the size of 18 × 24 mm2 and the thickness of 0.12 to 0.17 mm was employed in this study.

The attenuation characteristic of the borosilicate glass for the radioactive ray was evaluated by using a commercially available Geiger counter (LND-712 Geiger counter kit, inFlow Corp, Japan). A radiation source is a commercially available mantle�No.M-7909, PERL METAL Co. Ltd.�, which contains thorium as radioactive material. Figure 3 shows the radiation source of the mantle. When the radiation source was placed 20 mm away from the detecting window of the commercial Geiger counter, the attenuation characteristic of the radiation dose was evaluated. The schematic of the examination is shown in Fig. 4.

To evaluate the influence of the glass thickness, the thickness was varied by changing the number of the glass stacks. Figure 5 shows the result of the radiation dose as the function of glass thickness. The beta ray and the gamma ray were detected because thorium has alpha decay and beta decay. Since the alpha ray is immediately blocked out by the glass, it was not detected by GM tube. The beta ray is attenuated exponentially [5] by a glass with the thickness of 0.15 to 2 mm and the beta ray is blocked out for glass thickness with more than 2 mm. Because the gamma ray can not be blocked out for glass thickness with 6 mm, only gamma ray is detected at the glass thickness of 2 to 6 mm.

When the glass with the thickness of 150 �m was used, the detected radiation dose was 30 % lower than without glasses. It is confirmed that a thinner borosilicate glass window is preferable for the window of MEMS-GM tube.

25 mm

Fig.3 Radiation source has Thorium decay in this study.

Borosilicate glass

Geiger tube(LDN-712)

Radiationsource

Fig.4 Examination apparatus of attenuation property for borosilicate glass.

01234567

0 1 2 3 4 5 6Glass thickness (mm)

Rad

iatio

n do

se (�

Sv/h�

(Beta ray and gamma ray)

(Only gamma ray)

Fig.5 Measured attenuation property of borosilicate glass.

IV. FABRICATION OF MEMS-GM TUBE

A. Process flow Figure 6 shows a fabrication process flow of the MEMS-GM

tube. (1) Silicon wafer and pyrex glass wafer with holes that access to the sensor anode are bonded by anodic bonding.

(2)Photolithography process for silicon etching is carried out. (3)The silicon is etched through by Deep-RIE (Reactive ion etching) to form a cavity to encapsulate the inactive gas. After silicon etching, the photo resist is removed by O2 plasma ashing. (4) At last, top glass is bonded with the silicon/glass wafer using the polymer, KMPR (KMPR-1035, MICROCHEM ltd), epoxy resin. Then the inactive gas is confined into the cavity. Figure 7 shows a photograph of the silicon wafer during the process. The schematic of bonding method is shown in Fig. 8. The bonding conditions are load weight at 0.98 MPa and substrate temperature at 150 °C.

6767

(1)Anodic bonding (2)Lithography (3)Deep-RIE

(4)Bonding

B. Bonding Technology An inactive gas is required to be sealed into the cavity for

GM tubes. In MEMS technology, typical methods for bonding silicon and glass are polymer bonding and anodic bonding [6-9]. Thermoplastic resin is used for the polymer bonding. On the other hand, the anodic bonding supplies high voltage (400 to 1000 V) at high temperature (250 to 400 °C) to bond them. Because the anodic bonding makes covalent bond strongly between glass and Si molecules, the gas does not leak from the cavity. Since the general GM tubes operate at voltage range of 500 to 700 V, the anodic bonding is not suitable to seal the gas into the cavity under low pressure because of discharging. Hence, the anodic bonding is used to the first step bonding of bottom glass and silicon.

The second step polymer bonding is used for bonding with the top glass and silicon of the device. An epoxy resin can be adopted as an adhesive material of polymer bonding. A KMPR based on the epoxy resin is a negative type photoresist and can

be easily coated on the wafers by spin coating. However, several tens of �m step that called edge bead is formed at the edge of the wafer because of high viscosity of the KMPR. The KMPR film except for the edge part was exposed to avoid the forming of the edge bead. In addition, the soft bake time for the KMPR lithography was optimized to planarize the film. Table I. is shown lithograph condition of KMPR. Fig. 9 shows surface profiles of the KMPR spincoated glass wafer measured by white-light interferometer (New view 6300, Zygo corp). In Fig. 9(a), an edge bead is formed around the edge of the wafer. Removed the edge bead and reduction of the steps can be confirmed as shown in Fig. 9(b). A tensile test was carried out to investigate the bonding strength of the KMPR. The bonding strength of 4.7 MPa that is enough to use general application was obtained.

TABLE I.

LITHOGRAPHY CONDITION OF KMPR conventional improved

Spin coat

slope-5 sec� 500 rpm-10sec�

slope-10sec� 3000 rpm-30 sec

slope-5 sec� 500 rpm-10sec�

slope-10sec� 3000 rpm-30 sec

Soft bake

R.T.� slope-6 min�

65 °C-10 min� 100 °C-2 hour

R.T.� slope-6 min�

65 °C-10 min� 100 °C- 15 min

Exposure 50 sec 120 sec PEB 100 °C-3 min 100 °C-3 min

Development - 330 sec

V. DETECTION OF RADIOACTIVE MATERIALS

To evaluate detectable gas species and pressure, a hole was drilled at the bottom side of glass to adjust the gas pressure, and the hole was connected with a polyurethane tube. Figure 10 shows the photograph of the prototype MEMS-GM tube. (5 mm gap between electrodes of the device). Figure 11 shows the schematic diagram of the experimental system.

Fig.7 Test device after Deep-RIE.

(a) Surface profile for conventional KMPR coating condition.

Fig.9 Cross-section diagram of KMPR coated on 4 inch glass wafer.

(b) Surface profile for improved KMPR coating condition.

Distance (mm)

Hei

ght (

�m)

0 33.6 67.2 100.8 0

40

20

60

Distance (mm)

Hei

ght (

�m)

0 33.6 67.2 100.8 0

40

20

60

A A’ A’ A

A’A

A A’A A’ 50 �m

20 �m

Edge bead

surface of glass

Silicon

Glass

Photoresist

KMPR

15 mm

Silicon

KMPR Glass

Weight

Heater

Vacum chamber

Fig.8 Bonding method of KMPR.

Fig.6 Fabrication process of MEMS-GM counter tube.

Heater

Contanct to anode Contanct to cathode

2 cm

2 cm

6868

Figure 12 shows a circuit diagram of radioactive ray detection. High voltage was supplied between the anode and the cathode by using voltage source (TEOS5051, KIKUSUI products). Nitrogen was adopted for the inactive gas inside the cavity because it is easy to be got. The pressure of inactive gas in commercial GM tubes is about 10 to 27 kPa. Hence, the pressure was adjusted to 15.6 kPa in this study. The radiation source was approached and removed with voltage application of 2.6 kV to the MEMS-GM counter. The result is shown in Fig. 13. When the radiation source was approached to the MEMS-GM tube, the pulse derived from ionizing collision was observed. When the radiation source was removed away from the MEMS-GM tube, the pulse was not observed. At a result, it is clear that the fabricated MEMS-GM tube by using MEMS bonding technology has sensitivity to radioactive materials.

10 M��

��

F�

F�

MEMS-GM tube�5 kV�

0.5��

detector1 �1.2 k

Fig.12 Circuit for MEMS-Geiger counter.

0

2

4

6

8

10

12

0 2 4 6 8 10

Volta

ge(V

)

Time(s)

Approach to radiation source

Fig.13 Detected pulse.

Far from radiation source

In future work, to improve the sensitivity, the window of borosilicate glass will be changed to mica window, which has high radioactive transmittance. Also, optimum gas, pressure and sensor size are investigated.

Cathode Anode

15 mm

VI. CONCLUSION We proposed the MEMS-GM tube to miniaturize the

conventional Geiger counter. The MEMS-GM tube which has glass/silicon/glass stacked structure by using the MEMS bonding technology was fabricated. The borosilicate glass which attenuates the radiation dose to 70% was found to use the window of MEMS-GM tube. When the MEMS-GM tube using the borosilicate glass as the window was set up at the inactive-gas pressure of 15.6 kPa and supply voltage of 2.6 kV. It is confirmed that the MEMS-GM tube has a sensitivity for the radioactive source.

15 mm

Fig. 10 Prototype device.

Vaccum Mass flow gauge controller

REFERENCES [1] GLENN F. KNOLL, “RADIATUIN DETECTION AND

MEASUREMENT”, 1982, pp. 167–188. Valve

Rotary Pump

[2] D.H. Wilkinson, “The Geiger discharge revisited Part I. The charge generated”, Nucl.Instr.and Meth.A321,1992, pp.195-210.

[3] Herbert Fridman, "Geiger Counter Tubes”, Proceeding of the I.R.E Vol37,1949, pp791-808

[4] William J. Price, ”NUCLEAR RADIATION DETECTION”, 1958, pp122-152

N2

tank [5] GLENN F. KNOLL, “RADIATUIN DETECTION AND

MEASUREMENT”, 1982, pp. 33–72. [6] Thomas F.Marinis, Joseph W.Soucy, James G.Lawrence and Megam

M.Owens, ”Wafer Level Vacuum Packaging of MEMS sensors”,Proc Electron Compon Technol Conf Vol.55thVol2,2005,pp1081-1088

MEMS-GM tube

[7] Yong-Kook Kim, Eun-Kyung Kim, Soo-Won Kim and Byeong-K won Ju,”Low temperature epoxy bonding for wafer level MEMS packaging”, sensers and actuators A 143,2008,pp323-328

Fig.11 Setup for MEMS-GM tube.

[8] Martin A.Schemidt, "Wafer-to-Wafer Bonding for Microstructure Formation”, Proceedings of the IEEE Vol.86,1998,pp1575-1585

[9] Byeunhleul Lee, Seonho Seok and Kukjin Chun, ”A study on wafer level vacuum packaging for MEMS devices”, J.Micromech.Microeng.13, 2003, pp663-669

6969

�Abstract—To realize extremely low-power consuming sensor

systems, a SOHI (Silicon on Honeycomb Insulator) wafer, which has ultra-small parasitic capacitance, is proposed and an accelerometer is fabricated to demonstrate use of the technique. In case of using a low-power-consumption switched capacitor measurement circuit, parasitic capacitance decreases sensor sensitivity dramatically. The concept of the SOHI wafer is a special SOI (Silicon on Insulator) wafer with the buried SiO2 layer replaced with a honeycomb-shaped SiO2 thick layer (20 ��m thick) to reduce the parasitic capacitance. In this paper, two types of SOHI wafer, which could reduce the parasitic capacitance to about 21% and 2% of typical SOI wafers, respectively are proposed. On a trial production, the structure of a prototype three-dimensional accelerometer is fabricated. By demonstrating the application of sensor fabrication, the SOHI wafer is shown as a useable material as the base substrate. Also, the ability to realize the accelerometer with low parasitic capacitance by using the SOHI wafer is demonstrated.

Index Terms—Silicon on Insulator; Low parasitic capacitance; Honeycomb structure; Accelerometer

I. INTRODUCTION

YSTEMS for monitoring a human activity, such as heartbeat and movement, have attracted attention. The

authors have developed a human activity monitoring system by using integrated sensors and circuits [1],[2]. For monitoring human activities continuously over the long-term,low-power-consumption sensors and circuits are required. The authors have placed in the system a low power C/V converter circuit for driving the circuit and sensor, and for detecting acceleration. The C/V circuit dissipates approximately 11.3 �Wof electricity [3]. However, the parasitic capacitance of the

T. Yokomatsu is with the Maenaka Human Sensing Fusion Project, ERATO

Japan Science and Technology Agency (e-mail: [email protected]). H. Takeuchi is with the Maenaka Human Sensing Fusion Project, ERATO

Japan Science and Technology Agency (e-mail: [email protected]). Y. Miyagawa is with the Dept. of EECS Graduate School of Engineering,

University of Hyogo (e-mail: [email protected]).K. Kanda is with the Dept. of EECS Graduate School of Engineering,

University of Hyogo (e-mail: [email protected]). T. Fujita is with the Dept. of EECS Graduate School of Engineering,

University of Hyogo (e-mail: [email protected]). K. Higuchi is with the Maenaka Human Sensing Fusion Project, ERATO

Japan Science and Technology Agency (e-mail: [email protected]). K. Maenaka is with the Dept. of EECS Graduate School of Engineering,

University of Hyogo (e-mail: [email protected]).

sensor device has critical influence in the sensitivity when the C/V converter employed. The accelerometer needs to be produced with low parasitic capacitance to improve the sensitivity of the sensor. For fabricating a capacitive-type accelerometers (shown in Fig. 1), it is often used an Silicon on Insulator (SOI) wafers for the base substrate [4],[5]. While SOI wafer could make the fabrication process easier, it creates large parasitic capacitances between the fixed electrodes and the Si handle layer with much large values than the detection capacitance, as shown in Fig. 1 (b). For fabricating a sensor with low parasitic capacitance, the improvement of base substrate is required.

Novel Honeycomb SOI Structure with Low Parasitic Capacitance for Human-Sensing Accelerometer

Tokuji Yokomatsu, Haruka Takeuchi, and Yoshikazu Miyagawa, Non Member, IEEE

Kensuke Kanda, Takayuki Fujita, Kohei Higuchi, and Kazusuke Maenaka, Member, IEEE

S

Acceleration

Fixed Electrode

Fixed Electrode

A’

A

Spring SpringMovableElectrode

(a)

Fig. 1. Concept of accelerometer using SOI wafer: (a) image of top side, (b) cross section structure image of line A-A’

(b)

MovableElectrode

FixedElectrode

Detection Capacitor

ParasiticCapacitance

Device Layer

Insulation Layer

Handle Layer


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.21

70


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.21

70

In this paper, Silicon on Honeycomb Insulator (SOHI) wafer, which has thick insulation layer consisting of a 20 �m thick honeycomb insulation layer and a 1 �m thick SiO2 film, is proposed for use as the base substrate. Two types of SOHI wafer, which decrease the parasitic capacitance to 20% and 2% of standard SOI wafers, respectively, are proposed. By fabricating the structure of the accelerometer, the applicability of the SOHI wafer to MEMS process is verified, and ability of fabricating a low parasitic capacitance accelerometer is shown.

II. CONSIDERATION OF FABRICATION METHOD

To fabricate SOI wafer, silicon fusion-bonding is popular method[6]. To fabricate a low parasitic capacitance SOI wafer, it is effective to create a thick insulation layer before the bonding process is done. Several fabrication methods for making a thick insulation layer were considered, such as using SiO2 chemical vapor deposition (CVD) and thermal oxidation [7]-[11].

The CVD method can make thick insulation layer in lower temperature than thermal oxidation, and deposition speed can control by changing the amount of gas or pressure of atmosphere[7],[8]. However, quality of layer is lower than layer of oxidation, and adhesion strength for wafer bonding is unsure. Moreover, thick layer causes a large curvature. For these reasons, CVD method is not thought to be practical.

On the other hand, thermal oxidation method has takes a long time to obtain several �m of thickness of an insulation layer using the standard method. However, the oxidation time can be reduced to that of 1 �m thick oxidation by patterning many trench patterns with narrow pitches in the Silicon wafer before oxidation [9]-[11]. But this method has a problem of inducing large residual stress and making a curvature in the wafer. The residual stress has a negative impact to other fabrication processes.

Therefore, we proposed to form a hexagonal hole array and patterned a honeycomb structure to one of two wafers. Due to the hexagonal honeycomb structure, the wafer was resistant to residual stress, and was thought to reduce the curvature [12]. Furthermore, the thermal insulation properties of SOI wafer can be improved when fusion-bonding process is done in a vacuum. In this paper, two types of SOHI wafer are proposed: the honeycomb layer made by Si (named SOHI type 1), or made by SiO2 (named SOHI type 2).

III. DESIGN OF SOHI WAFER

Fig. 2 shows the concept of the SOHI. Fig. 3 is the schematic illustration of the cross-sectional views and equivalent circuits with parasitic capacitance of the SOI and SOHI wafers. A 1 mm2 area is considered for all three wafers.

The SOI wafer is assumed to have a 1 �m thick insulation layer, and the parasitic capacitance is calculated to be approximately 32 pF. On the other hand, in the SOHI wafer, the hexagonal honeycomb structure which consist of a hexagonal array of vacuum holes with 20 �m depth and a separating wall of 2 �m width would reduce much of the parasitic capacitance. The contact area between the handle and the device layers for

SOHI is about 20% of the SOI wafer with same area. To compare SOHI type 1 with SOI wafer, assuming the decrease in the contact area to 20%, parasitic capacitance is reduced toabout 21%, which is approximately 6.4 pF. Furthermore, to compare SOHI type 2 with SOI wafer, parasitic capacitance is reduced to about 2%, which is approximately 0.64 pF. The SOHI type 2 has not only a low parasitic capacitance, but also a feature to enhance the breakdown voltage and thermal insulation.

The SOHI wafer is intended to be fabricated with the typical MEMS process. By confirming the process method, the SOHI wafer can be used as the base substrate for sensor fabrication.

IV. FABRICATION OF SOHI WAFER

A. Fabrication of SOHI type 1 The fabrication process of SOHI wafer type 1 is indicated in

Fig. 4: fabricate the 1 �m thick SiO2 layer with thermal oxidation to the top Si wafer (a), pattering the honeycomb

SOI SOHI Type 1 SOHI Type 2

Appearance

Cross section

Equivalent circuit

SiSiO2 SiO2

SiO2Si

Si

Si

Si

Si

Si

SiO2SiO2 SiO2

Si

Si

Si

Si

Si

vacuum vacuum

(a) (b) (c)

Fig. 3. Comparation of SOI wafer and SOHI wafers. Each image shows an appearance, cross section and equivalent circuit of (a) SOI wafer, (b) SOHI wafer type1 and (c) SOHI wafer type 2

Fig. 2. Concept of SOHI wafer

2 �m

20 �m

7171

structure with 20 �m depth to bottom Si wafer with Deep Reactive Ion Etching (DRIE) (b), silicon fusion-bonding between two wafers in vacuum (c), and polishing the device layer to the intended thickness (d). Fig. 5 shows the Scanning Electron Microscope (SEM) image of the honeycomb structure after DRIE. The average width and depth of the pillar were measured to be 2.27 �m and 21.1 �m, respectively. By using a wafer-bonding machine, the two wafers were bonded with a load pressure of 2 MPa for 30 min under the 4×10-4 Pa atmospheric condition and a temperature of 400 ˚C. After bonding, the wafer was heated for 2 hours at the heater temperature of 1100˚C. Fig. 6 (a) shows the cross-sectional image after direct bonding. The image demonstrates a successful bonding and no breakage failures of the honeycomb structure due to pressure. Fig. 6 (b) shows the SEM images of bonded surface after peeling wafers out. The traces of the destroyed honeycomb walls indicate sufficient bonding

strength between the honeycomb structure and the oxide film surface. For the polishing process, a polishing machine (PM-6,

LOGITECH Corp.) was used. The process steps are as follows: (1) 415 �m was roughly polished in 2 hours with using a slurry of SiC#800, (2) next 50 �m was roughly polished in 1 hour with SiC#1500 slurry, (3) next, 20 �m was polished and it finally became the mirror surface in approximately 4 hours with a slurry of SiC#6000. The trial fabrication of the SOHI wafer with the thickness dimensions of the device layer, oxide film layer and the honeycomb structure of 50 �m, 1�m and 21.1 �m, respectively, was completed.

B. Fabrication of SOHI type 2 Ideally, the type 2 SOHI wafer can be fabricated by oxidized

the honeycomb structure according to the way in Fig. 4. However, this process method is difficult to fabricate because of occurrence of thermal expansion and the strong curvature of the wafer due to thermal oxidation.

Fig. 7 (a) shows an SEM image of oxidized honeycomb structure. The bonding surface of the honeycomb wall has become more round after the thermal oxidation. This reduces the contact area of wafer bonding and results in the failure of fusion-bonding.

Therefore, the fabrication process has been improved to make a SiO2 honeycomb wall with a flat surface. The process flow is shown in Fig. 8: patterning a hexagonal honeycomb trench of 20 �m depth with a DRIE (a), thermal oxidation for 8 hours at 1100 ˚C and fill all trench with SiO2 (b), polish the thermal

(a)

(c)

(d)(b)Fig. 4. Process flow of SOHI. (a) Thermal oxidation of the top wafer. (b) Etching honeycomb structure to bottom wafer. (c) Direct bonding of two wafers. (d) Polishing of top Si to make device layer.

SiO2

SiO2

Si

Si

(a) (b)Fig. 5. SEM images of Honeycomb structure. (a) Image of top side wafer. (b) Cross section image.

10 �m10 �m

2.27��m21 �m

20 �m 20 �m

(a) (b)

Fig. 6. SEM images of SOHI wafer after direct bonding. (a) Cross section image. (b) Image after peeling away the device wafer from handle wafer .

(a)

(b)

(c)

(d)

(e)

(f)Fig. 8. Improved fabrication process of SOHI. (a) Patterning a hexagonal honeycomb trench with a DRIE. (b) Thermal oxidation to honeycomb patterned wafer and filled a trench with SiO2 .(c) Polishing a thermal oxidized wafer to make a flat side.(d) Etching a Si surface to appear the SiO2 honeycomb wall. (e) Direct bonding of the two wafers. (f) Polishing of top Si to make device layer.

Si

Si

SiO2

Fig.7. SEM images of SiO2 honeycomb pattern.(a)After the normaloxidation. (b) After etching the Si cavity in Fig.8.

(b)(a)

10 �m 10 �m

7272

oxidized wafer for 2 �m to shave the SiO2 surface and make a flat side (c), etch a Si surface for 20 �m with DRIE to reveal the SiO2 honeycomb wall (d), silicon fusion-bonding between two wafers in vacuum (e), and polish the top Si of the wafer to make device layer (f). Figure 7(b) shows an SEM image of honeycomb wall after etching the silicon surface. Comparing Fig. 7(b) with Fig. 7(a), it can be seen that the rounder shape of the honeycomb wall has become flatter.

On the other hand, the improved method for cancelling the strong curvature perfectly has yet to be confirmed. However, fabricating the honeycomb structure to the both sides of the bottom wafer is thought to an effective method. By fabricating the oxidized honeycomb structure on both sides of bottom wafer, the curvature radius of bottom wafer, which is calculated to be approximately 5 m, has been increased to approximately 30 m. Since the curvature radius of a silicon wafer is approximately 50 m, this method is useful for curvature improvement.

The completion of SOHI type 2 will be reached soon.

V. FABRICATION OF ACCELEROMETER

In order to verify the applicability of the SOHI wafer to the MEMS process, various type of accelerometer structure were fabricated using an SOHI wafer type 1. The process flow is shown in Fig. 9: Preparing the SOHI wafer type 1 (a), forming electrode pads with Au/Cr metal (b), patterning the design of the accelerometer to the device layer with DRIE (c), and etching the SiO2 layer under the movable electrode and spring with BHF wet etching process (d). Fig. 10 is one image of the accelerometer structures . A central mass is supported by four springs, and the applied acceleration moves the mass. There are four fixed electrodes with comb shapes separated from the mass by electrostatic gaps. The fixed electrode structures are anchors. The sensor is designed to detect acceleration by a capacitance change of the comb electrode.

For SiO2 layer etching, 16 BHF (made by Morita chemical industries Corp., LTD in Japan) was used. From the data sheets, the etching speed of SiO2 is written as 0.1 �m/min in temperature of 22 . When it used for SOI wafer, side etching rate of SiO2 layer has become 0.1 �m/min. In the case of the SOHI wafer, however, the side-etching rate would change because of honeycomb structure. Fig. 11 shows an etching flow of SiO2 layer; etched a SiO2 layer in a vertical direction as much as thickness of layer (a), filled hexagonal hole of honeycomb structure with BHF and started etching the joint area of honeycomb wall and SiO2 layer (b), completed the etching of joint area and suddenly fill the next hole with BHF (3), repeatsthe flow of (2) and (3). Since the etching step is quite different from those of SOI wafer, it is necessary to check the side etching rate.

For trial, structure of accelerometer has dipped in BHF for 60minutes. After it, fixed electrode was peeled off, and SiO2 layer under the fixed layer was observed. Fig. 12 shows an SEM image of the accelerometer after SiO2 layer etching. Fig. 12(a) shows a SEM image that the fixed electrodes has peeled off,and Fig. 12(b) shows the close up image of honeycomb structure under the fixed electrode. In Fig. 12(b), honeycomb structure which could see in white is area of not etched. From Fig. 12(a), amounts of side etching is confirmed as much area as of three hexagonal holes of honeycomb structure. Since the etching speed of SiO2, thickness of SiO2 layer, designed width of honeycomb wall and pitch of wall were 0.1 �m/min, 1 �m, 2�m and 20 �m, respectively, average speed of side-etching is calculated to 1 �m/min.

Fig. 10 . Image of Accelerometer

Comb Electrode

Spring

Mass

(a)

(b) (d)

(c)Au/Cr

Fig. 9. Process flow of the accelerometer. (a) Preparing the SOHI.(b) Forming an electrode pad with Au/Cr. (c) Etching a pattern of accelerometer with DRIE. (d) Etching a SiO2 layer with BHF wet etching.

SiSiO2

Si

100 �m 10 �m

Finished Etching

SiO2 layer yet etched

(a) (b)Fig. 12. Side etching of SiO2 layer. (a) peeling off the fixed electrode. (b) Close up image of honeycomb structure

(a)

(b)

(c)

(d)Fig. 11. Flow of SiO2 layer etching. (a) Etching in a vertical direction. (b) Filled the honeycomb hole with BHF and started etching the joint area of honeycomb wall and SiO2 layer. (c) Completed to etching a joint area and filled the BHF with anotherhole. (d) Repeat the flow(c).

SiSiO2

Si

BHF

7373

To make sure that the SiO2 layer is adequately etched and the moveable structure could move smoothly without sticking, the resonant frequency of the accelerometer structure was measured. For the evaluation, a laser Doppler vibrometer (MLD-103A, NEOARK Corp.) was used. While the designed value of the accelerometer shown in Fig. 10 was analyzed to be 12.9 kHz by using finite element analysis software ANSYS,measurement result was observed to be 9.8 kHz, as shown in Fig. 13. Furthermore, another type of accelerometer was measured a sensitivity of the differential capacitance with a commercialized C/V converting circuit. While the designed value was 3 fF/G, measurement result was 1.5 fF/G. The decrease in the resonant frequency and sensitivity would be attributed to the thinner spring width and the expand comb gap width derived from photolithography and etching process.

As a result, an accelerometer structure, which can vibrate with nearly the designed value, was fabricated on SOHI wafer. From the fabrication, SOHI wafer has shown to be used as the base substrate for sensor fabrication.

VI. CONCLUSION To fabricate an accelerometer with low parasitic capacitance

and high sensitivity, the SOHI wafer with thick insulated layer consisting of a 20 �m thick honeycomb structure and 1 �mthick oxide film layer was proposed and fabricated. By using SOHI type 1, the parasitic capacitance derived from the insulation layer was reduce to about 21% of the value of a conventional SOI wafer. Furthermore, SOHI type 2 which is nearly finished being realized can reduce parasitic capacitance to about 2% of a conventional SOI wafer with the improvements of enhanced breakdown voltage and thermal insulation. In addition, the structure of accelerometer was fabricated on SOHI wafer type 1 in trial. In spite of the difference of etching rate of the SiO2 layer, the structure was fabricated with a typical MEMS process. The resonant frequency of the structure was measured to be 9.8 kHz, which is close to the design value.

As a result, we show that SOHI wafer can be used as the base substrate for sensor fabrication. Furthermore, we show the possibility to realize an accelerometer with low parasitic capacitance by using an SOHI wafer.

REFERENCES

[1] Kazusuke Maenaka“Human Sensing Fusion Project for Safety And Health Society” IEEJ Trans. SM Vol. 128, Issue 11, pp. 419-422, 2010 (In Japanese )

[2] Kazusuke Maenaka, et al., “Human Sensing Fusion – For Health and Safety Life”, Proc. 5th Asian-Pacific Conf., Transducers and Micro-Nano technl., p.36, 2006

[3] Yoshikazu Miyagawa, Tokuji Yokomatsu, Haruka Takeuchi, Kensuke Kanda, Takayuki Fujita, Kohei Higuchi and Kazusuke Maenaka,”Acceleration sensor Using SOI wafer with Honeycomb Insulator”, Proc. IEEE SMC 2012, to be published

[4] Henrik Rodjegard, et al., “A monolithic three-axis SOI-accelerometer with uniform sensitivity”, Sensors and Actuators A, vol. 123-124 (2005), pp.50-53

[5] Hiroki Kuwano, “Fundamentals and applications of MEMS/NEMS Engineering”,Technosystem Co, first eddition(2009), pp.589-591, in Japan

[6] Tai-Ran Hsu,“MEMS and MICROSYSTEMS”, Second edition, JOHN WILEY & SONS, INC.,

[7] J.W. Klaus and S. M. George,“SiO2 Chemical Vapor Deposition at Room Temperature Using SiCl4 and H2Owith an NH3 Catalyst”, Journal of the Electrochemical Society, vol .147, No. 7, pp.2658-2664 (2000)

[8] S. Wolf and R.N.Tauber, “Silicon Processing for the VLSI Era, Vol. 1: Process Technology”, Lattice Press, Sunset Beach, CA (1986)

[9] Kensuke Kanda, Kyota Mitsuyama, Jun Nakamura, Takayuki Fujita, Kohei Higuchi, Kazusuke Maenaka, “Silicon dioxide as Mechanical Structure Realized by Using Trench-Etching and Oxidation Process”, Proc. Eurosensors XXV, Procedia Engineering, vol 25, pp.843-846, 2011

[10] Chunbo Zhang and Khalil Najafi,“Fabrication of thick silicon dioxide layers for thermal isolation”,JOURNAL OF MICROMECHANICS AND MICROENGINEERING, vol. 14(2004), pp.769-774

[11] G. Barillaro, A. Diligenti, A. Nannini and G. Pennelli, “A thick silicon dioxide fabrication process based on electrochemical trenching of silicon”, Sensors and Actuators A, vol. 107 (2003), pp.279-284

[12] A.-J. Wang and D.L. McDowell, “Yield surfaces of various periodicmetal honeycombs at intermediate relative density”, International Journal of Plasticity, vol. 21 (2005), pp.285–320

0 5 10 15 20

Am

plitu

de (a

.u)

1.2

1.0

0.8

0.6

0.4

0.2

0

Fig. 13. Frequency characteristic of accelerometer

Frequency (kHz)

9.8 kHz

7474

Flexible Sensor for Human Monitoring System by Using P(VDF/TrFE) Thin Film

Takayuki Fujita, Shohei Shiono, Kensuke Kanda, Kazusuke Maenaka

Dept. of EECS, Graduate School of Engineering University of Hyogo

Himeji, Hyogo, Japan [email protected]

Hiroyuki Hamada, Kohei Higuchi Maenaka Human Sensing Fusion Project

ERATO, Japan Science and Technology Agency Himeji, Hyogo, Japan

Abstract—Continuous monitoring of human physical and mental conditions are heavily required solving issues to realize a high quality of life society. The human condition data especially; heartbeat, ECG (electrocardiogram), respiration and blood pressure are very important vital signs to prevent a human sudden disease, such as a heart attack. In this study, we fabricated a novel sensor to monitor the heartbeat and respiration, which is made on a flexible thin film of P(VDF/TrFE) as a sensing material and PDMS (poly dimethyl siloxane) as a substrate material. The PDVF (poly vinylidene fluoride) that is thin, flexible and piezoelectric material can be used for the flexible force sensor. The fabrication steps and some experimental results for human monitoring system are described.

Keywords-Human monitoring, flexible sensor, PVDF, piezoelectric, thin-film

I. INTRODUCTION The accidents in our daily life are occurred frequently,

which depends on healthiness and wellness of individuals. In order to enhance the quality of our daily life, continuous and noninvasive monitoring systems for healthcare are strongly required to realize [1-3]. In previous work, our paper presents a flexible force sensor for human monitoring systems by using P(VDF/TrFE) and PDMS (poly dimethyl siloxane) material [4]. It will realize the system that continuously monitor the human condition data such as; heartbeat, ECG, respiration and blood pressure. Especially, the heartbeat is very useful for prediction of the human sudden accident. There are large varieties of monitoring systems for the heartbeat measurement, however we don't have definitive device.

The aim of this study is to develop a novel wearable human monitoring system (Fig. 1). The system will consist of various MEMS (micro electromechanical systems) sensors, microcontroller (MCU), application specific IC

(ASIC), RF module and power source on a flexible substrate [3]. In this study, we try to fabricate a flexible heartbeat and a respiration sensor device in the flexible substrate. The proposed device consists of the flexible and stretchable electric wiring on a PDMS substrate, a PVDF force sensor and also an integrated measurement circuitry. The substrate can easily attach on the human skins by adhesiveness of PDMS. A MEMS compatible batch fabrication process and experimental results are also described.

II. FLEXIBLE SUBSTRATE FOR FLEXIBLE ELECTRIC WIRING WITH BIOCOMPATIBILITY

The PDMS is one of the most popular biocompatible materials for human application. It can use for fabrication of a fine structure by a molding technique. The PDMS also have a high adhesiveness by adjusting the mixing ratio of the linear- and cyclo-PDMS. The PDMS with mixing ratio of 10:0.2 shows very good adhesiveness that can keep stick to human skins during a week [5]. Regarding the PDMS substrate as a printed circuit board (PCB), the mixing ratio of 10:1 is good for hardness. The flexible substrate for attached on the human skins is combination of the sticky and hard PDMS layers of 0.5 mm and 1.0 mm thickness, respectively. Fig. 2 shows the fabricated test piece of the combined PDMS substrate with adhesive layer and adhesion test on the human wrist.

An Au/Cr thin film electrode can be sputtered and micro patterned on the PDMS that acts as the PCB substrate. However the sputtered electrode easily break when the PDMS substrate is bended or stretched. In order to improve a robustness of the thin film electrodes, we proposed roughened surface substrate. Stress of the electrode film will be reduced when the film substrate is wavy roughened as shown in Fig. 3. The roughened PDMS is fabricated by using the PDMS molding on the roughened glass substrate that is grinded by sandblasting. After that the release agent polymer was deposited on the glass surface. The roughness is

Sensors�Power supply�

MCU/ ASIC� RF� Antenna�

Flexible substrate�Heart beat, respiration sensor�

Figure 1. Pastable human monitoring system.

Figure 2. Sticking test (left) and adhesion test during a week on the

subject’s wrist.


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.22

75


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.22

75

controlled by the condition of the sandblasting, such as time, pressure and particle size of blasting. Fig. 4 shows the structure of test specimens to evaluate the flexible electrode. The type-1 has no surface roughening. The type-2 and type-3 have an average surface roughness, peak-to-peak bump level of 0.28 µm, 2 µm and 1.90 µm, 12 µm, respectively.

In order to evaluate robustness of the flexible electrode on the PDMS substrate, it has four tests of (A) tensile test, (B) repeat tensile test, (C) bending test, and (D) thickness dependence of electrode.

A. Tensile test Fig. 5 shows electrode resistance versus maximum

deformation ratios before the electrodes breaking. The size of Au/Cr electrode with thickness of 300/30 nm is 1 × 30 mm2 that is buried in the PDMS substrate with size of 0.3 × 10 × 60 mm3. The type 1 specimen breaks immediately when a deformation is applied. Type 2 and 3 show good robustness for 20% and 30% of deformation ratio. More wavy and rougher surface has a good robustness to the tensile stress, but higher resistance.

B. Repeat tensile test Since the type-3 specimen showed good robustness, a

repeat tensile test is applied to type-3 one. Fig. 6 shows the electrode resistance versus repeat number of tensile test with deformation ratios of 5% (dashed line in Fig. 6) and 10% (solid line in Fig. 6). The tensile force was applied alternately. The data show the tensile and relaxed resistance of the electrode. The resistance increase slightly, but the electrode can work over 100,000 and 5,000 cycles for 5% and 10% deformation ratio, respectively.

C. Bending test Three types of specimens are tested with bending test.

The specimens are looped around the circular cylinder jig. Fig. 7 shows the resistance of electrodes versus the curvature radius of the cylinder for three types specimens. The specimens with surface roughening show considerably good robustness for bending than non-roughening one.

D. Thickness dependence of electrode In order to evaluate an optimum electrode thickness, the

tensile test with various electrode thicknesses was performed. Fig. 8 shows the result of the breaking deformation ratio versus Au electrode thickness. An applied current for conduction test is 1 µA, and a tension rate of 100 µm/sec is applied to the type-3 specimen. The breaking ratio of the applied deformation is inverse proportion to the Au electrode thickness. Thus the thin electrode shows high robustness for the tensile force, but it has low capacity of conduction current. It is tradeoff.

0

10

20

30

40

50

� 100 67 28 16 12Diameter[mm]

Resistance[Ω]

Type1Type2Type3

271Ω

Curvature radius [mm]

Res

ista

nce

[Ω]

��50 33.5 14 8 6

Figure 7. Bending test results for three types of specimen.

0

20

40

60

80

100

120

140

1 10 100 1000 10000 100000Cycle

Resistance[Ω]

Tension(5%)Release(5%)Tension(10%)Release(10%)

Number [times]

Res

ista

nce

[Ω]

Figure 6. Tensile test of 5 and 10% deformation ratio for type-3.

0

5

10

15

20

25

30

0 10 20 30 40Strain[%]

Resistance[Ω]

Type1Type2Type3Break�

Break� Break�

Deformation ratio [%]

Res

ista

nce

[Ω]

Figure 5. Result of the tensile test.

�Type 1

�Type 2

�Type 3

PP level ( 2 µm )

Ra=13.7 nm

PP level ( 12 µm)

PDMS(300 mm) Au(300 nm)/ Cr(30 nm)�

PDMS(300 mm)

Cross section�

Ra=0.279 µm

Ra=1.897 µm

Surface photograph�

Figure 4. Test specimen structures with rough surface.

PDMS substrate

Electrode film

Figure 3. Flexible electrode on the roughen PDMS

7676

III. CHARACTERIZATION OF P(VDF/TRFE) FILM The piezoelectric materials that generate a charge from a

mechanical stress are good candidate for human monitoring sensors because it has possibilities to realize an ultra low power or non-power consumption sensor. There are some sorts of piezoelectric material, such as PZT (lead zirconate titanate), AlN, PVDF, etc. Of these, the PVDF thin film is very suitable material for human sensing device because the PVDF film is a polymer with very high piezoelectric voltage coefficient of g and has flexibility and batch fabrication compatibility. As PVDF and its copolymers also have remarkable mechanical and chemical properties, they have been widely used for sensor, actuator, and energy harvesting device applications [6-8].

One of the PVDF copolymer of P(VDF/TrFE) can easily fabricate with spin coating method. Since the composition of VDF/TrFE is very effective to achieve high piezoelectric coefficient and low current leakage, an optimized P(VDF/TrFE) copolymer powder (KFW#2200, KUREHA Corp. Japan) with a composition of VDF/TrFE = 75/25 mol% is used for our experiments. At first, the powder is dissolved in MEK (methyl ethyl ketone) at 60 ºC to obtain a solution of 10 wt.% and then spin coated on the Au surface of the silicon substrate at 1700 rpm. After drying and annealing the sample at 140 ºC for 2 hours in N2 gas ambient, P(VDF/TrFE) film with a thickness of 2 µm is achieved. The sample surface is sputter deposited with aluminum layer with a thickness of 500 nm as the upper electrodes.

Fig. 9 shows the 2θ X-ray diffraction pattern of the P(VDF/TrFE) film. A high diffraction peak at about 2θ = 20° shows a characteristics of piezoelectric crystal. The (110) or (200) diffraction of P(VDF/TrFE) is observed

In order to evaluate piezoelectric characteristics of the P(VDF/TrFE) thin film, hysteresis curves of polarization and electric field (P-E curve) were measured by using hysteresis measurement equipment (TF Analyzer 2000, aixACCT Systems GmbH., Germany). Fig. 10 shows the P-E hysteresis curves of the P(VDF/TrFE) test sample, which is 2 µm thickness and the electrode area of 0.855 mm2. The remanent polarization and coercive electric field are found to be as high as 71.7 mC/m2 and 73.1 MV/m, respectively. The film shows good ferroelectricity to realize force sensors.

IV. FABRICATION OF HEARTBEAT AND RESPIRATION SENSOR

The P(VDF/TrFE) film in MEMS processing, it will be required a capability to the fine patterned fabrication technique. In previous work, we established the dry etching process of the film [4]. Process gas O2 flow-rate is 50 sccm, and process pressure of 5 Pa. The etching rate is 300 nm/min at an antenna power of 200 W and a platen power of 20 W. By using Al electrode film as an etching mask, the minimum processing accuracy is 10 µm.

Fig. 11 shows the fabrication process flow of the flexible heartbeat and respiration sensor that we proposed. (a) Roughened surface PDMS with thickness of 300 µm is prepared as same as type-3 test sample. (b) Bottom electrode that is made of Au/Cr with thickness of 300/30 nm is deposited and patterned. (c) P(VDF/TrFE) solution is spin coated and receives thermal treatment. After annealing, the film thickness is 25 µm. (d) Al for upper electrode is deposited and patterned, then the P(VDF/TrFE) film is dry etched with Al mask. (e) Wire bonding for upper and lower electrodes. (f) Finally, the cover PDMS layer with the adhesive layer is fabricated.

Photographs of the sensor during the fabrication (top) and after the wire bonding (bottom) are shown in Fig. 12. The thicknesses of PDMS and P(VDF/TrFE) are 300 µm and 25 µm, respectively. The sensor size is 12 × 36 mm2 and capacitance is about 3.1 nF.

-100-80-60-40-20

020406080

100

-120 -60 0 60 120Electric field [MV/m]

Pola

riza

tion

[mC

/m 2 ]

kPo

lariz

atio

n [m

C/m

2 ]�

Applied electric field [MV/m]� Figure 10. Measured P-E hysteresis curve of the patterned

P(VDF/TrFE) film with a thickness of 2 µm.

0

5

10

15

20

0 0.5 1 1.5

Au�[µm]

��[%]

Au thickness [µm]

Def

orm

atio

n ra

tio a

t bre

akin

g [%

]

Figure 8. Thickness dependence of electrode.

1

10

100

5 10 15 20 25 30 352θ [deg]

Inten

sity

[cou

nts]

g

P(VDF/TrFE) (110) or (200)

Inte

nsity�

��theta �º�� Figure 9. XRD pattern for P(VDF/TrFE) film.

7777

V. EXPERIMENTAL RESULT OF FLEXIBLE SENSOR Fig. 13 shows the schematic diagram of the measurement

apparatus for the heartbeat and the respiration sensing. The flexible sensor is pasted on the subject’s body (male, 24 years old, coauthor, on left side of his chest). The output signal from the sensor is amplified 1000-fold and filtered by a 10 Hz LPF. The measured waveform from the flexible sensor is shown in Fig. 14. The graph also shows ECG data that simultaneously measured by another sensor. By using the flexible sensor, the heartbeat of 114 BPM (beat per min.) and respiration of 12 times per min. are clearly obtained. Since the flexible sensor measures chest movement, the heartbeat and respiration data are mixed output. Thus, when the subject holds his breath, the data show only heartbeat waveform (Fig. 15).

VI. FUTURE WORKS In this state, the PDMS substrate is just used for

protection cover and foundation of the P(VDF/TrFE) film. Now we fabricate an integrated flexible sensor by using PDMS film as a PCB substrate. Fig. 16 shows the fully integrated flexible sensor that consists of P(VDF/TrFE) flexible sensor and a peripheral circuitry are assembled on the PDMS substrate. Discreet electric parts on the Au/Cr film wiring construct the circuitry by using soldering. The area of the circuitry is 18 × 10 mm2 and entire size of the

-2

-1

0

1

2

3

4

0 2 4 6 8 10

Time[sec]

Out

put V

olta

ge[V

] k

Out

put v

olta

ge [1

V/d

iv]�

Time [sec]

Heartbeat�

ECG

Figure 15. Measurement data when subject hold subject’s breath (after

amplification and LPF).

-2

-1

0

1

2

3

4

0 2 4 6 8 10Time[sec]

Out

put V

olta

ge[V

] k

Heartbeat�

Respiration�

Time [sec] O

utpu

t vol

tage

[1 V

/div

]� ECG

Figure 14. Measurement data of heartbeat and respiration from flexible sensor (after amplification and LPF). ECG data is used as a reference.

ECG

×1000

10 Hz LPF

Output +

Figure 13. Schematic diagram of measurement apparatus for heartbeat

and respiration sensing.

(a)�

(b)�

(c)�

(d)�

(e)�

(f)�

Figure 11. Fabrication process flow of flexible sensor.

Bottom electrode (Au/Cr) Top electrode (Al)

36 mm

12 mm

Figure 12. Photographs of flexible sensor.

7878

integrated sensor is 12 × 46 m2. The signal noise ratio will be improved by on-device instrumentation pre-amplifier.

VII. CONCLUSION In this study, the flexible heartbeat and respiration sensor

for human monitoring system was demonstrated. The sensor has a flexible and stretchable electrode by using roughened surface of PDMS substrate material. The electrode with thickness of 300 nm on the roughened PDMS with bump level of 12 µm shows good robustness for both bending and stretching. The sensor part was made of P(VDF/TrFE) piezoelectric polymer material, which has high piezoelectric coefficient and flexible characteristics. The P(VDF/TrFE) that laminated by the PDMS with flexible electrode constructed the flexible sensor. Some experimental results for the human heartbeat and respiration were successfully measured.

REFERENCES [1] T. Hara, Y. Matsumura, M. Yamamoto, T. Kitado, H. Nakao, H.

Nakao, T. Suzuki, T. Yoshikawa, and S. Fujimoto, “The Relationship Between Body Weight Reduction and Intensity of Daily Physical Activities Assessed with 3-Dimension Accelerometer,” Jpn. J. Phys. Fitness Sports Med., Vol.55, No.4, pp. 385–392, 2006.

[2] A. DeHennis and K. D. Wise, “A Wireless Microsystem for the Remote Sensing of Pressure, Temperature, and Relative Humidity,” J. Microelectromechanical Systems, Vol.14, No.1, pp. 12–22, 2005.

[3] K. Maenaka, T. Fujita, H. Hamada, X.C. Hao, Y. Iga, Y.G. Jiang, K. Kanda, K. Kuramoto, M. Saito, Q. Wulan, D. Zhu and K. Higuchi, “Human Sensing Fusion -For Health and Safety Life-”, 5th Asia-Pacific Conference on Transducers and Micro-Nano Technology, p. 36

[4] Y. G. Jiang, H. Hamada, S. Shiono, K. Kanda, T. Fujita, K. Higuchi and K. Maenaka,“A PVDF-based flexible cardiorespiratory sensor with independently optimized sensitivity to heartbeat and respiration”, Procedia Engineering , 2010, 5, pp. 1466-1469

[5] H. Hamada, Y. G. Jiang, K. Kanda, T. Fujita, K. Higuchi and K. Maenaka, "Development of a real time monitoring system combining a human body patch type cardiography sensor with RF transceiver", Proc. of The 5th Asia-Pacific Conference on Transducers and Micro-Nano Technology (APCOT2010), p. 261

[6] S. Choi and Z.A. Jiang, "Novel wearable sensor device with conductive fabric and PVDF film for monitoring cardiorespiratory signals", Sensors and Actuators, 2006; A128, pp.317-326.

[7] P. Finkel, S.E. Lofland and Ed. Garrity, "Magnetoelastic/piezoelectric laminated structures for tunable remote contactless magnetic sensing and energy harvesting", Appl. Phys. Lett., 2009, 94, 072502

[8] Y.G. Jiang, S. Shiono, H. Hamada, T. Fujita, K. Higuchi and K. Maenaka, "Low-frequency energy harvesting using a laminated PVDF cantilever with a magnetic mass", Proc. PowerMEMS 2010, pp. 375-378

46 mm

12 m

m

Figure 16. Integrated flexible sensor device for heartbeat and

respiration monitoring.

7979

Feature Definition using Dependency Relationsbetween Terms for Improving Nursing-care Text

Classification

Manabu Nii∗†, Yoshinori Hirohata∗, Atsuko Uchinuno‡ and Reiko Sakashita‡∗ Graduate School of Engineering, University of Hyogo, Himeji, Hyogo, Japan

Email: [email protected]† WPI Immunology Frontier Research Center, Osaka University, Suita, Osaka, Japan‡ College of Nursing Art and Science, University of Hyogo, Akashi, Hyogo, Japan

Abstract—In order to improve the nursing-care quality, “Webbased Nursing-care Quality Improvement System” have beenproposed and operating continuously. In the proposed system,for evaluating actual nursing-care process, freestyle Japanesetexts which are called “nursing-care texts” are collected throughthe Internet in Japan. The nursing-care experts can evaluateactual nursing-care process and recommend some improvementsto nurses by reading the collected nursing-care texts carefully.Since the number of nursing-care experts who can evaluatethe nursing-care texts is a few, it is hard to do the abovementioned evaluation process for a large number of nurses. Toassist nursing-care experts in evaluating the nursing-care texts, acomputer aided nursing-care text classification system has beendeveloped. In this paper, we propose a method to improve theclassification performance of the computer aided nursing-caretext classification system. Dependency relation between terms isextracted from the nursing-care text and used the dependency asa feature value which represents characteristics of the nursing-care text.

I. INTRODUCTION

The assessment criteria for the nursing-care quality were

proposed in Japan to improve the nursing-care quality. For-

merly, the nursing-care experts have visited hospitals and

observed nurses and their actual nursing-care, according to the

defined criteria. In such an evaluation procedure, a nursing-

care expert can observe only one or a few nurses at a time.

For a few nursing-care experts, it is difficult to evaluate the

actual nursing-care which are performed by a large number

of nurses in Japan by this evaluation procedure. Since the

actual nursing-care observation is difficult for a small number

of experts, a freestyle nursing-care text collection system

via the Internet has already been developed. The freestyle

nursing-care texts contain answers which are written based

on the nursing-care performed by nurses actually. Therefore,

the nursing-care experts can evaluate actual nursing-care of

nurses. After evaluating such texts, the experts will recommend

some improvements to nurses by reading the collected nursing-

care texts carefully. However, it is still hard work to do the

above mentioned evaluation process for a large number of

nurses. In order to help nursing-care experts in evaluating

the collected nursing-care texts, a computer aided nursing-

care text classification system has been developed [1]–[10].

Although a fuzzy classification system [1] and a neural net-

work based classification [2] have been proposed, currently

the computer aided nursing-care text classification system

uses support vector machines (SVMs) [3]–[10]. In the above

mentioned system, since the nursing-care texts classified by

nursing-care experts are used as the training data, the trained

SVMs as well as experts can classify nursing-care texts.

In our previous works [3]–[9], feature values are defined as

the existence of terms, and in [10] we have proposed a new

feature definition using dependency relations between terms.

While dependency relations are very important for us, the fea-

ture definition based on existence of terms lacked dependency

relation information between terms to understand texts. There-

fore, we need a novel feature vector making method based on

dependency relations between terms. In this paper, to improve

the classification performance of our SVM based classification

system, dependency relations between terms is extracted from

the nursing-care text and used the dependency relations as

feature values which represent characteristics of the nursing-

care text. Moreover, we propose a feature definition which

considers compound nouns in the dependency relations.

II. COMPUTER AIDED NURSING-CARE TEXT EVALUATION

In this section, our Web based nursing-care quality improve-

ment system is briefly described. Then, the computer aided

nursing-care text classification system and its components

which have been proposed in [3]–[10] are also presented.

A. Overview of the Web based nursing-care quality evaluationsystem

Due to the insufficient number of nursing-care experts, the

Web based system has been developed. A set of ID and pass-

word to access the Web site are assigned to every nurses who

join our nursing-care quality improvement program. Using

such ID and password, each nurse logs in to the Web site.

In the Web site, several questions which are according to

the defined assessment criteria for the nursing-care quality

are presented to nurses. For answering the questions, nurses

write freestyle texts which represent their own nursing-care,

and then submit them to the Web site. Of course, because every


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.68

110


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.68

110

nurses are placed on different situation, some nurses may not

be able to answer some questions. In such case, the nurses may

leave empty the answers. This is because the answers which

are written by nurses based on the actual nursing-care are

required to evaluate the nursing-care quality. The nursing-care

texts are sets of such answers for a wide variety of questions

and include information of their own patients’ condition and

technicality of their actual care, etc. Such nursing-care texts

are gathered per each question as a nursing-care text set via

Web site and stored into our database.

The nursing-care text classification is processed as follows.

First, collected nursing-care texts are decomposed into mor-

phemes by using morphological analysis software. A term

list per each nursing-care text set is generated by using

decomposed nouns and verbs. Then, each text is converted

into a numerical vector(i.e., a feature vector). We can choose

the following methods which are described in the following

subsections to tackle to several problems, if needed. For the

lack of terms, we can use the “retrieving term information from

Web”. For considering synonyms, “generating dictionaries of

conceptual fuzzy sets” can be used. Or for finding terms sets

which contribute to classification, “GA based term selection”

methods can be used.

B. Feature vector making based on the term existence

We need to make numerical vectors from the nursing-care

texts for constructing our classification system. In Japanese

texts, terms are not separated by a space. First, a text is

converted to a term vector by decomposing it into morphemes.

The morphological analysis software “MeCab [11]” is used for

this purpose. All nouns and verbs in decomposed morphemes

are stored into a term list and the other terms are not used in

the current classification system. TLQiyear denotes the term list

for the question Qi, which stores terms collected in the year.

Every nursing-care text is converted into a feature vector by

assigning ’1’ to each feature when the corresponding term

to the feature in the nursing-care text existed in TLyear,

and otherwise ’0’ is assigned to the feature. Additionally, the

following three features are joined to the binary feature vector.

Nmorph,p is the number of all morphemes in the text p. Nterm,p

is the number of terms (i.e., nouns and verbs) in the text p.

tfidfsum,p is total tfidf value of p as follows:

tfidfsum,p =∑

ti∈ptfi,p · idfti (1)

tfi,p =Nti,p,p

maxNk,p(2)

idfti = logNtexts

dfti(3)

Step 1 A nursing-care text is selected and decomposed into

some morphemes.

Step 2 Nouns and verbs are extracted from the decomposed

morphemes.

Step 3 If an extracted term is existed in the term list TLyear,

then ‘1’ is assigned. Otherwise ‘0’ is assigned. And

additional three features (i.e., Nmorph,p, Nterm,p, and

tfidfsum,p) are joined. As a result, a feature vector is

expressed as (4).

xp = (Nmorph,p, Nterm,p, tfidfsum,p, b1, . . . , bNTLyear)

(4)

where NTLyearmeans the number of terms in the term

list TLyear.

Step 4 If all texts in the nursing-care text set have been

converted, then finish this procedure. Otherwise go to

Step 1.

Our classification systems are constructed by using the feature

vectors which have been already evaluated by nursing-care

experts as a training data set.

C. GA based term selection

We have already reported in [4] that the classification

performance is improved when specific terms for each question

Qi are appropriately chosen. In [4], the specific terms were

selected by hand according to some indexes. However, we

have no knowledge about which term or combination of

terms improve the classification performance. Therefore, it

is difficult to select the specific terms or combinations of

terms from all term list. For selecting the specific terms or

combinations of terms automatically, GA based term selection

method have been proposed in [5], [6]. The term selection

problem was defined as a combinatorial optimization with two

objectives. The GA based term selection method can select

terms or combinations of terms for improving the classification

performance of the classification system. In [5], [6], the op-

timization problem was defined as two-objective optimization

which has (1) maximizing the number of correctly classified

texts, and (2) minimizing the number of selected terms. The

first objective is the most important and the second objective

is also important because of understandability of the results.

For optimizing the two objectives, the weighted sum of the

two objectives was used as the objective function of GA as

follows:

Maximize f(s) = w1 ·NCP (s)− w2 ·NST (s), (5)

where s is an individual defined in (6)-(7) (see also the

equation(4)), NCP (s) is the number of correctly classified

patterns (i.e., nursing-care texts), and NST (s) is the number

of selected terms.

s = (c1, c2, . . . , cNTLyear) (6)

cl =

⎧⎨

⎩

1 : if the term which is corresponding to clis used for classification,

0 : otherwise.

(7)

D. Retrieving term information from Web

In practical use of the nursing-care classification system,

the classification system will be constructed by using the

last year’s data set. And then the current year’s texts will

111111

be classified by the constructed system. In this situation, a

problem due to the difference of the number of terms between

TLlast and TLcurrent is caused. Terms which are appeared

in the TLcurrent for the first time are unknown terms for the

trained classification system. That is, a nursing-care text which

has many unknown terms is represented as a feature vector

which consists of many ‘0’. It is difficult for the classification

system to classify such feature vectors.

One of techniques to tackle the above mentioned problem

is to use known terms in the term list which are most similar

to the unknown terms. In order to consider that the unknown

terms in the new data set are approximately the same as the

known terms, a technique of using information from Web is

proposed [8].

E. Conceptual Fuzzy Sets based Dictionary

The method of retrieving term information from Web men-

tioned in the previous subsection is effective for the completely

unknown terms. However, since the method retrieves on unse-

lected Web sites, the search results contain some results which

are unsuitable for our purpose (i.e., nursing-care area).

As another idea for unknown terms, we have proposed a

conceptual fuzzy set based dictionary in [9]. A conceptual

set is defined as a category of terms in the conceptual fuzzy

set based method. For example, a conceptual set “pain”

includes “hard pain”, “ache”, and “pain from a cancer”, etc.

Meaning “pain” has many expressions (i.e., Kanji compound)

in Japanese. An application of using the conceptual fuzzy sets

has been proposed in [12]. An atomic fuzzy set consists of a

prototype and the relative terms. A set of atomic fuzzy sets is

defined as a dictionary of conceptual sets. In order to construct

a dictionary of conceptual sets, fuzzy c-means method was

used. Corpora are clustered by a fuzzy c-means method by

using cosine similarity.

III. FEATURE DEFINITION USING DEPENDENCY RELATIONS

In our previous research, existence of each term in a

text was used as the numerical feature vector. The feature

vector xp (see, (4)) is basically a binary vector. Such a

binary feature vector does not include the dependency relations

between terms. The dependency between terms is important

for understanding texts. Therefore, a feature definition which

considers the dependency relations is needed for expressing

characteristics of texts.

In this section, we propose a new derivation for generating

feature vectors based on [10]. For understanding Japanese or

any other texts, the dependency analysis plays an important

role as a basic process [13]. “Cabocha” which was proposed in

[13] is one of the Japanese dependency analysis software. The

authors reported that Cabocha achieved about 90% accuracy

for parsing Japanese dependency. Therefore, we use it for

finding dependency in our nursing-care text because Cabocha

has the highest accuracy among Japanese dependency parsers.

We can find dependency relation information by using

Cabocha (e.g., see Fig. 1). For example, the ellipse #1 has

a dependency relation to the ellipse #2 in Fig. 1. Since we

�

�

Fig. 1. An example of dependency relation by Cabocha.

have focused on nouns and verbs in our previous research, the

dependency among nouns and verbs is considered in this paper.

All nouns and verbs are extracted from the training data set

and stored into a term list. This procedure is the same as our

previous method. Then, each text is processed by Cabocha for

finding dependency. Cabocha can find indexes of dependency

relation of “from terms” and “to terms”.

In Fig. 1 example, the #1 is structured by two nouns. In

such case, we consider two types of feature definitions.

1) Don’t care of compound nouns: When a set of terms

A = ai, i = 1, . . . , nA, has a dependency relation to the other

set of terms B = bj , j = 1, . . . , nB , we consider that each

element of A has dependency relation to the set of terms B. If

the set B consists of some terms, we define the feature value

as follows: Each element of A has dependency relation to the

last term of B. The feature value of the elements of A is the

index of the last term of B. Figure 2 shows an illustration of

a case of “don’t care of compound nouns” feature definition.

2) Consideration of compound nouns: On the other hand,

when a set of terms A = ai has a dependency relation to the

other set of terms B = bj , we consider that the last element

112112

Fig. 2. An example of Don’t care of compound nouns.

of A (i.e., the last term of A) has dependency relation to the

set of terms B. In this consideration, the terms a1, ...., anA−1

have the dependency relations between ai and ai+1. That is ,

ai has a dependency relation to ai+1, and the last element anA

has a dependency relation to the last element of B. Therefore,

the feature value of the elements of ai is the index of ai+1

and the feature value of the last element anAis the index of

bnB. Figure 3 shows an illustration of a case of “consideration

of compound nouns” feature definition.


In this section, experimental results on the actual nursing-

care texts are shown. To evaluate the effectiveness of our new

feature definition, we omit the first three features which are

used in our previous works (Nmorph,p, Nterm,p, and tfidfsum,p

in (4)). And also “retrieving term information from Web”,

“generating dictionaries of conceptual fuzzy sets”, and “GA

based term selection” methods were not applied. We just com-

pared two types of our feature definition, which is based on

dependency relation between terms, with our previous feature

definition based on existence of terms. Table I summarizes of

the nursing-care text sets using in this paper. The training data

sets were made by using 2007 and 2008 text sets. The SVM

based classification systems were trained by the training data

sets. Then the nursing-care text sets of 2009 were classified

by the trained SVMs. Tables II and III show the classification

results of our proposed method “Don’t care of composed

nouns”. And Tables IV and V show the classification results

of our proposed method “Consideration of composed nouns”.

We can see from the tables that our both new feature definition

achieved about 10% – 20% higher classification performance

than the previous definition. When we consider compound

Fig. 3. An example of Consideration of compound nouns.

TABLE ISUMMARY OF NURSING-CARE DATA SET.

Question ID The number of texts

2007 2008 2009

P123 548 569 806

P131 500 519 754

P132 439 400 626

P212 493 481 700

P213 509 509 720

P221 483 512 755

P222 455 483 691

P322 441 407 593

P411 392 374 561

P425 318 442 680

P431 446 431 652

P611 375 355 567

P613 440 429 662

P621 449 432 652

nouns, the classification performance was almost the same

when using our proposed feature definition of “don’t care of

compound nouns”. Therefore, we can conclude that the depen-

dency relation based feature definition is effective to classify

the nursing-care texts. It seems natural that composed nouns

were considered when we utilize the dependency relations for

making feature vectors. However, from experimental results,

we have almost the same performance using both methods.

We need to further consideration about these two definitions.

113113

TABLE IIRESULTS OF THE PROPOSED METHOD OF DON’T CARE OF COMPOSED

NOUNS FOR 2009 TEXT CLASSIFICATION USING 2007 TRAINING DATA.

proposed(DC of CN) conventional

Question ID Rate(%) (NCP / all texts) Rate(%) (NCP / all texts)

Q123 60 482 / 806 42 338 / 806

Q131 45 339 / 754 42 319 / 754

Q132 33 206 / 626 22 139 / 626

Q212 68 477 / 700 48 338 / 700

Q213 44 318 / 720 40 285 / 720

Q221 72 541 / 755 45 340 / 755

Q222 81 559 / 691 57 397 / 691

Q322 61 363 / 593 46 273 / 593

Q411 91 509 / 561 63 354 / 561

Q423 88 594 / 672 52 348 / 672

Q425 62 424 / 680 32 218 / 680

Q431 89 579 / 652 58 376 / 652

Q611 72 406 / 567 47 267 / 567

Q613 38 251 / 662 36 240 / 662

Q621 84 545 / 652 56 362 / 652

TABLE IIIRESULTS OF THE PROPOSED METHOD OF DON’T CARE OF COMPOSED


proposed(DC of CN) conventional


Q123 59 477 / 806 56 449 / 806

Q131 53 400 / 754 26 194 / 754

Q132 35 221 / 626 22 140 / 626

Q212 57 398 / 700 45 316 / 700

Q213 45 321 / 720 33 240 / 720

Q221 71 539 / 755 24 180 / 755

Q222 81 559 / 691 47 327 / 691

Q322 61 364 / 593 48 282 / 593

Q411 90 503 / 561 58 327 / 561

Q423 89 596 / 672 52 348 / 672

Q425 62 424 / 680 34 233 / 680

Q431 88 576 / 652 60 394 / 652

Q611 73 412 / 567 54 305 / 567

Q613 60 399 / 662 38 252 / 662

Q621 84 545 / 652 31 203 / 652

V. CONCLUSION

In this paper, two types of feature vector making methods

which is based on the dependency relations were proposed

to improve the classification performance of the computer

aided nursing-care text classification system. The proposed

feature vector making methods are based on a novel feature

definition which is based on dependency relation between

terms. Classification results show higher classification perfor-

mance than our previous feature definition which was based

on the existence of terms. While the previous definition just

includes one kind of information about term existence, the new

feature definition proposed in this paper includes both term

existence and dependency relation between terms. Moreover,

TABLE IVRESULTS OF THE PROPOSED METHOD OF CONSIDERATION OF COMPOSED


proposed(consideration of CN) conventional


Q123 60 486 / 806 42 338 / 806

Q131 45 338 / 754 42 319 / 754

Q132 33 207 / 626 22 139 / 626

Q212 57 402 / 700 48 338 / 700

Q213 45 321 / 720 40 285 / 720

Q221 71 539 / 755 45 340 / 755

Q222 72 499 / 691 57 397 / 691

Q322 60 357 / 593 46 273 / 593

Q411 88 496 / 561 63 354 / 561

Q423 89 599 / 672 52 348 / 672

Q425 62 424 / 680 32 218 / 680

Q431 88 574 / 652 58 376 / 652

Q611 71 405 / 567 47 267 / 567

Q613 38 249 / 662 36 240 / 662

Q621 84 545 / 652 56 362 / 652

TABLE VRESULTS OF THE PROPOSED METHOD OF CONSIDERATION OF COMPOSED


proposed(consideration of CN) conventional


Q123 60 486 / 806 56 449 / 806

Q131 53 400 / 754 26 194 / 754

Q132 35 221 / 626 22 140 / 626

Q212 57 401 / 700 45 316 / 700

Q213 45 321 / 720 33 240 / 720

Q221 71 539 / 755 24 180 / 755

Q222 72 499 / 691 47 327 / 691

Q322 60 357 / 593 48 282 / 593

Q411 88 496 / 561 58 327 / 561

Q423 89 599 / 672 52 348 / 672

Q425 62 424 / 680 34 233 / 680

Q431 88 574 / 652 60 394 / 652

Q611 71 405 / 567 54 305 / 567

Q613 60 399 / 662 38 252 / 662

Q621 84 545 / 652 31 203 / 652

we considered compound nouns to define feature values. By

using our proposed feature definition, we observed higher

classification performance over almost all data sets.

REFERENCES

[1] T. Sakaguchi, M. Nii, Y. Takahashi, A. Uchinuno, and R. Sakashita,“Nursing-care Data Classification using Fuzzy Systems”, Proc. of SCIS& ISIS 2006, Tokyo, Japan, pp. 2017–2022.

[2] M. Nii, Y. Takahashi, A. Uchinuno, and R. Sakashita, “Nursing-careData Classification Using Neural Networks”, Proc. of 2007 IEEE ICMEInternational Conference on Complex Medical Engineering, Beijing,China, pp. 431–436.

[3] M. Nii, S. Ando, Y. Takahashi, A. Uchinuno, and R. Sakashita, “Nursing-care Freestyle Texts Classification using Support Vector Machines”,Proc. of 2007 IEEE International Conference on Granular Computing,CA, pp. 665–668.

114114

[4] M. Nii, S. Ando, Y. Takahashi, A. Uchinuno, and R. Sakashita, “Featureextraction from nursing-care texts for classification”, Proc. of 6th Inter-national Forum on Multimedia and Image Processing, 2008, in CDROM(6pages).

[5] M. Nii, S. Ando, Y. Takahashi, A. Uchinuno, and R. Sakashita, “GAbased Feature Selection for Nursing-Care Freestyle Text Classification”,Proc. of Joint 4th International Conference on Soft Computing andIntelligent Systems and 9th International Symposium on advancedIntelligent Systems, 2008, pp. 756–761.

[6] M. Nii, T. Yamaguchi, Y. Takahashi, R. Sakashita and A. Uchinuno,“Analysis of Nursing-care Freestyle Japanese Text Classification UsingGA-based Term Selection”, Proc. of World Automation Congress 2010,Kobe, Japan, IFMIP495 (CD-ROM, 6 pages).

[7] M. Nii, T. Yamaguchi, Y. Takahashi, A. Uchinuno, R. Sakashita, “Im-proving Classification Performance of Nursing-Care Text ClassificationSystem by Using GA-based Term Selection”, Journal of AdvancedComputational Intelligence and Intelligent Information, Vol.14, No.2,pp. 142–149.

[8] M. Nii, T. Yamaguchi, Y. Mori, Y. Takahashi, R. Sakashita, andA. Uchinuno, “Classification of Nursing-care Data using AdditionalTerm Information”, Proc. of SCIS & ISIS 2010, Okayama, Japan,pp. 1469–1474.

[9] M. Nii, Y. Takahashi, A. Uchinuno, and R. Sakashita, “An Approachusing Conceptual Fuzzy Sets for Nursing-care Text Classification”, Proc.of 2012 World Automation Congress, Puerto Vallarta, Mexico, 2012,WAC 2012 1569535307.

[10] M. Nii, Y. Hirohata, A. Uchinuno, and R. Sakashita, “New FeatureDefinition for Improvement of Nursing-care Text Classification”, The2012 IEEE International Conference on Systems, Man, and Cybernetics(IEEE SMC 2012), (Accepted).

[11] MeCab, Yet Another Part-of-Speech and Morphological Analyzer,http://mecab.sourceforge.jp/.

[12] T. Nomura, K. Yamauchi, and T. Takagi, “Context Aware Query Re-finement System using conceptual fuzzy sets and Generation MethodEmploying Fuzzy Clustering”, Japan Society for Fuzzy Theory andIntelligent Informatics, 2009, Vol. 21, No. 2, pp. 175-183.(in Japanese)

[13] T. Kudo, Y. Matsumoto, “Japanese Dependency Analysis using Cas-caded Chunking”, CoNLL 2002: Proceedings of the 6th Conferenceon Natural Language Learning 2002 (COLING 2002 Post-ConferenceWorkshops), 2002, pp. 63–69.

115115

�Abstract— A tool for measuring the quality of nursing care was developed in Japan in 1993. This tool consists of three dimensions: structure, process, and outcome. One aspect of the outcome dimension, specifically nursing-sensitive medical incidents such as falls, downfalls, pressure ulcers, hospital infections and drug-related errors was measured. In this study, aspects of nursing care that may influence the frequency of medical incidents were examined. As the above five medical incidents were differentially related to each other, they were summarized under two factors using principal component analysis. Factor 1 comprised falls, downfall, pressure ulcers and drug-related errors. Factor 2 comprised hospital infections and pressure ulcers. All domains of the structure and process of "incident prevention" were related to factor 1. These variables were able to explain 46.1% of the variance in factor 1. Most of all the domains of structure, and process of "patient empowerment", "direct care", and "incident prevention" were related to factor 2. These variables were able to explain 45.5% of the variance in factor 2.

Index Terms—quality of nursing care, medical incident, quality improvement, website

I. INTRODUCTIONN 1989, the first step towards the development of a measurement system for evaluating the quality of nursing care was taken by Dr. H. Minami and her research group. The

group developed two types of questionnaire, one for patients and the other for nurses. Subsequently, a measurement tool to evaluate and improve the quality of nursing care was developed between 1993 and 1997. A Delphi study was conducted to determine the components associated with quality of nursing care[1][2]. Subsequently, five qualitative studies were conducted to identify the nursing skills associated with quality of nursing care and to extract six domains of this nursing care[3]-[8]. The resultant tool was composed of three dimensions, namely Structure, Process, and Outcome.

Each dimension was constructed by measurement of the six domains, which were "understanding individuality", "patient empowerment", "family care", "direct care", "medical team coordination", and "incident prevention" as shown in Fig. 1.

Questions related to the structure items were expected to be answered by the head nurse and those related to the process

items by staff nurses. As outcome indicators, patient and family satisfaction and the frequency of medical incidents were assessed (Fig. 2).

The aim of this study was to clarify, using this QI database[9], the type of structures and nursing processes that could affect nursing-sensitive medical incidents.

Fig.1 Framework of nursing care system

Nursing Quality Related to Medical Incidents

Reiko Sakashita1, Atsuko Uchinuno1, Kazuko Kamiizumi2, Keiko Tei2 and Masumi Murakami2�1 College of Nursing Art and Science, University of Hyogo

College of Nursing, Aomori University of Health and Welfare, [email protected]

I

Process bystaff nurses

O u tc o m eR e s u lt b y U n it

QI Team

Structure by the head nurse

G e n e ra l in fo rm a tio n

R e s u lt o f H o sp ita lP ro c e s s

S tric tu re

Outcome bypatients & family+Number of incidents

Through web site, data is collected.

Improvement

consultation

Recommendation& intervention

�

��

��

��

��

� ��

R e s u lt b y U n itAnalysis

Nursing Care Quality Improvement System on Website

Fig.2 Outline of the web system


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.69

116


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.69

116

II. DATA AND METHODSA total of 145 Units, 552 staff nurses and 5987 patients

(2854 males and 2750 females) were included in this study in 2008.�Analysis was done by unit base, with score by staff nurses and patients averaged by unit.

The following five categories of nursing-sensitive medical incidents: falls, downfall, pressure ulcers, hospital infections, and drug-related errors, were measured by the unit base and reported by the chief nurse as an outcome indicator via the website.

The frequency of medical incidents was analyzed descriptively. Spearman's correlations were performed to analyze relationships within the five categories of incidents, and between incidents and nursing items. Incidents were summarized into factors using principal component analysis.

Using regression analysis, predictors (items) for incident factors were estimated, and synthetic variables from items of each domain that predicted incident factors well were calculated. Using those synthetic variables, structural equation modeling (SEM) was performed using Amos ver.17 [10] [11].

III. RESULT

A. Frequency of medical incidents

The frequency of medical incidents is shown in Fig. 3 and TABLE 1. Two large outliers, suspected to be due to reporter miscalculation, were observed. They were excluded from subsequent analysis.

TABLE 1 Frequency of medical incidents per 1000beds

N=145 B. Relationships among medical incidents

The five categories of medical incidents were differentially related to each other as shown in TABLE 2. Falls were significantly correlated with downfall and drug-related errors (p<0.05). Pressure ulcers were significantly correlated with hospital infections (P<0.01).

TABLE 2 Spearman's correlation among medical incidents.

�Downfall

Pressureulcers

Hospital infections

Drug-related errors

Falls .192* .039 .013 .180*

Downfall .163 -.003 .109Pressure ulcers .288** .012

Hospital infections

-.006

* p<0.05, **p<0.001 N=143

Principal component analysis was conducted to summarize these incidents. They were summarized under two factors.Factor 1 included falls, downfall, pressure ulcers and drug-related errors. Factor 2 included mainly hospital infections and pressure ulcers. These two factors were able to explain 68.2% of all variance of the variable. Distribution of these factors is shown in Fig. 4. Using these factors as dependant variables, we investigated what type of nursing quality was related to medical incidents.

TABLE 3 Factor coefficients of medical incidents.

Incidents mean s.d. medium min. max.Falls 1.89 1.88 1.60 0.0 16.4 Downfall 0.98 2.72 0.38 0.0 27.0 Pressure ulcers 0.94 3.17 0.39 0.0 32.8 Hospital infections 0.58 1.09 0.00 0.0 9.1 Drug-related errors 2.84 4.07 2.13 0.0 32.8

� Factor1 Factor2 Falls .823 -.122Downfall .614 -.131Pressure ulcers .786 .341Hospital infections .022 .967Drug-related errors .796 -.137

Fig. 3 Distribution of incidents per 1000 beds.

N N

N N

N

117117

C. Nursing items related to medical incidents

Spearman's correlations between incident factors and nursing items are shown in TABLE 4. Bed utilization was found to be preventive for factor 1, and also showed a positive relationship with facto2.

TABLE 4 Spearman's correlations between incident factors and nursing items.

Factor1 correlationcoefficient

Bed utilization rate at hospital -.215**

Bed utilization rate at unit -.246**

S212:Regular check of pamphlets -.221**

S213:Organization of information materials for patients

-.293**

S223:System to make discharge plan with patient -.246**

S412:Regular check of nursing standards -.227**

S414: Frequent reconsideration of nursing protocol

-.204*

S421:Materials for soap bed bath .166*

S551: Nurses work together in a good atmosphere -.187*

S611:Asesment system to prevent pressure ulcers -.266**

P633: Adequate staff to prevent incidents -.251**

Factor 2

Bed utilization rate at hospital .248**

Bed utilization rate at unit .221**

Assignment of pharmacist at unit .258**

S313:Enough space for family at the unit -.215*

S314:Another space for family to talk with patient -.343**

S521:System for change in staff shifts .206*

S522:System to fill a vacancy .188*

p132:Discuss role with patient -.179*

p223: Making discharge plan with patient -.177*

p321:Understanding family burden -.190*

p621:Confirming if the medical order is thought to be wrong.

-.246**

-(minus) means prevents incidents. N=143

D. Prediction of incident factors Regulation analysis was performed to check what aspects of

nursing quality could prevent incidents. We selected effective items for predicting incident factors using regulation analysis.

All structure domains (S1-S6), process domains (P1-P6) and utilization of beds at the unit were put into regression analysis to predict factor 1. Overall, nine variables were effective predictors of factor 1 (TABLE 5). All domains of structure and process of incident prevention (P6) were related to factor 1. These nine variables showed high coefficients of determination (R2), and were able to explain 46.1% of the variance of factor 1.

All structure domains (S1-S6), process domains (P1-P6), utilization of beds at the unit and the assignment of pharmacists at the unit were put into regression analysis to predict factor 2. Overall, eight variables remained as effective predictors of factor 2 (TABLE 6). Most of the domains of structure, processof "patient empowerment", "direct care", and "incident prevention" were related to factor 2. These variables showed

high coefficients of determination (R2), and were able to explain 45.5% of the variance of factor 2.

TABLE 5 Results of regression analysis to predict factor 1. � Standardized B t value p value intercept � -2.080 .039S1 -.196 -3.035 .003S2 -.180 -2.607 .010S3 -.258 -3.943 .000S4 -.154 -2.303 .023S5 -.172 -2.634 .009S6 -.244 -3.683 .000P4 -.140 -2.079 .040P5 -.124 -1.891 .061P6 -.206 -3.127 .002F=12.65 p<0.000 R=0.679 R2=0.461 N=143

TABLE 6 Results of regression analysis to predict factor 2. � Standardized B t value p value Intercept � -.021 .983S1 -.123 -1.877 .063S3 -.272 -4.194 .000S4 -.160 -2.386 .018S5 -.297 -4.462 .000S6 -.172 -2.605 .010P2 -.273 -4.149 .000P4 -.136 -2.050 .042P6 -.205 -3.140 .002F=13.89 p=0.000 R=0.675 R2=0.455 N=143

E. Explanatory model for incident factors

Using synthetic variables prepared through regression analysis, SEM was performed using Amos ver.17.The initial model is shown in Fig. 5. From the initial model, a better model was explored, using model fitness indicators such as GFI and AIC.

To explain factor 1, synthetic variables S1f1-S6f1, P1f1-P6f1, and bed utilization rate at the unit were put into the analysis. The final model is shown in Fig. 6, which showed a good fit to the data (GFI=0.954, AGFI=0.928). Every domain of Structure had a strong effect on prevention on factor 1 incidents. Process also had an impact on factor 1 incident prevention. In contrast, bed utilization rate did not play a major role.

To explain factor 2, synthetic variables S1f2-S6f2, P1f2-P6f2, bed utilization rate at the unit and the number of pharmacists at the unit were put into the analysis. The final model is shown in Fig 7, which shows a good fit to the data (GFI=0.957, AGFI=0.925). Both structure and process had a marked impact on prevention of factor 2 incidents, which were mainly those related to hospital infection and pressure ulcers. Compared with these domains, bed utilization rate at the unit and assignment of pharmacists at the unit did not play major roles.

118118

Fig. 5 Initial model of SEM

Standardized coefficients were shown.

Fig. 6 Final model explaining Factor 1 by SEM

Standardized coefficients were shown.

Fig. 7 Final model explaining Factor 2 by SEM

IV. CONCLUSIONS

As these five medical indicators were differentially related to each other, there might be common causes of these incidents. In this study, the importance of nursing quality for preventing medical incidents was confirmed. Almost half of the variance in incidents could be explained by nursing quality items. Thus this web-based nursing care quality improvement system makes it possible for nurses to reflect on and improve their care, to collect a very large quantity of data for building the Japanese national standardized evaluation system, identify factors for

better outcomes, and provide evidence for good practice in the future.

ACKNOWLEDGMENT The authors would like to express their sincere appreciation

to the head nurses, staff nurses and patients who participated in the study. This study was supported in part by Japan Society for the Promotion of Science with Grant-in-Aid for Scientific Research (B) (KAKENHI 18390572).

REFERENCES

[1] Uchinuno A. et al. Dimensions of Nursing Care Quality: Through Use of Delphi Method, The Japanese Journal of Nursing Research 1994:27: 315-323

[2] Chikazawa N. et al. Research Trends on Assurance of Nursing Care Quality: Literature Review, The Japanese Journal of Nursing Research 1994:27: 324-333

[3] Katada N. Qualitative Analysis of Nursing Care Skills, The Japanese Journal of Nursing Research 1996:29: 2-4

[4] Katada N. et al. Nursing Care Skills for Pain Alleviation, The Japanese Journal of Nursing Research 1996:29: 5-21

[5] Uchinuno A. et al. Nursing Skills for Monitoring Patients, The Japanese Journal of Nursing Research 1996:29: 23-33

[6] Yamamoto A. et al. Nursing Skills for Family Care, The Japanese Journal of Nursing Research 1996:29: 35-46

[7] Takezaki K. et al. Nursing Assessment and Skills in Promoting Self-Care Activities of the Patients, The Japanese Journal of Nursing Research 1996:29: 47-57

[8] Chikazawa N. et al. Nurses` Skills for Creating Interdisciplinary Cooperative Environment, The Japanese Journal of Nursing Research 1996:29: 59-70

[9] Kamiizumi K. History of developing nursing quality improvement system in Japan. The Japanese Journal of Nursing Research 2010: 43(5), 373-376.

[10] Joredkog K. G. Structural analysis of covariance and correlation matrices. Psychology 1978:27(14),1277-1301.

[11] Arbuckle J.A. Amos User's Guide Version 3.6. Chicago. Small Waters Corporation. 1997

119119

� Abstract— This article describes a personal identification

method using dynamic foot pressure change while walking. This system acquires foot pressure change by wearable pressure sensor. The sensor has four sensing tips for the each foot, and these tips are fixed on shoes. In the experiment, we acquire eight steps pressure change data. The system extracts gait features from every step. The features are based on peak pressure value of each sensing tip. For all steps and all registered subjects, we calculate Euclidean distance between the feature and template of subject. The template is made from learning data of each registered person. For each step, the system chooses a registered person with the shortest Euclidean distance as a candidate of walking person. Then, the system counts selected number of the steps of a person. Finally, we identify the walking person by the larger number. The number less than threshold, identification of this walking person is failure. We employed 10 volunteers and identify them. In the experiment, we took pressure data 10 times for each volunteer. We used 3 data for learning and used the other 7 data for test data. The proposed method obtained 0.83% in FRR and 0.02% in FAR, when threshold parameter less than 3.

Index Terms— personal identification, pressure senor, walking, wearable sensor.

I. INTRODUCTION ERSONAL identifications are widely used for many

situations at home or office. For example, the systems are used for health monitoring, record of the ingress / egress room, security and so on. This type personal identification requires high user-friendliness, because it is used at living space and they are used mundanely. Behavioral features of biometrics [1]-[4] such as signature or walking [5]-[17] identify person with high user-friendliness. The walking never force user to train, because it is the most natural motion in the behavioral features. In addition, by using walking, we are able to identify moving person without stopping their walking. Our identification system use walking as behavioral feature. Generally, a camera usage is popular to obtain walking data.

Manuscript received May 31, 2012. T. Takeda is with Graduate School of Engineering, University of Hyogo,

Hyogo, and Himeji Initiative in Computational Medical and Health Technology, Hyogo, Japan (corresponding author to provide phone: +81-79-267-4986; e-mail: [email protected]).

K. Kuramoto, Syoji Kobashi and Y. Hata are with Graduate School of Engineering, University of Hyogo, Hyogo, Japan, WPI Immunology Frontier Research Center, Osaka University, Osaka, Japan and Himeji Initiative in Computational Medical and Health Technology, Hyogo, Japan

However, the obtained images are strongly influenced by illumination or other environmental conditions. Moreover, people have psychological resistance for the camera because they do not like to watch by it.

In our study, we propose a biometric personal identification while walking, which employs dynamic foot pressure change. We have proposed several identification methods using mat type load distribution sensor [11]-[14]. However, these mat type load distribution sensors obtain walking data at limited space while walking. Moreover, they can acquire only two or three steps. In this paper, we employ wearable pressure sensor to obtain walking data anywhere. This sensor obtains many steps. The wearable pressure sensor is fixed on shoes. When user walks in the shoes, we are able to acquire dynamic pressure change data, and use it for identification. For reducing cost of the pressure sensor, we optimize the channel and employ four channels pressure sensor for the each foot. We fix four channels on right shoe and other four channels on left shoe.

We first acquire pressure change while walking by wearable pressure sensor. Secondly, we extract gait features based on peak of pressure change from each step pressure data. Thirdly, we calculate Euclidean distance between these features and templates for all steps. We identify subject by these Euclidean distances. Fourth, we show experimental results on 10 volunteers. We compare the proposed method with an identification method using foot shaped pressure sensor sheet. Finally, we conclude the technical results.

II. PRELIMINARIES

A. Wearable pressure sensor We acquire foot pressure change by wearable pressure

sensor. Fig. 1 illustrates overview of data acquisition system. As shown in Fig. 1, the system consists of eight channels pressure sensor (Nitta Corp., Octsens), a control equipment and personal computer. Circular tips are sensing points. Diameter of the circular tip is 9.5 mm. The control equipment converts the electrical resistance of each sensing tip into 8 bits (256 levels) digital value, and provides these pressure values to a personal computer. The sampling frequency is 100 Hz.

B. Data acquisition First, we fix the pressure sensor on shoes. It is necessary to

fix sensor on the optimal positions. As shown in Fig. 2, we fix circular tips of the pressure sensor on heel (PtA), outside of arch (PtB and PtC) and toe (PtD). Fig.3 shows an example of the fixed

Biometrics Personal Identification by Wearable Pressure Sensor

Takahiro Takeda, Student Member, IEEE, Kei Kuramoto, Member, IEEE, Syoji Kobashi, Member, IEEE, and Yutaka Hata Fellow, IEEE

P


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.70

120


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.70

120

pressure sensor. Second, a subject wears the shoes for acquisition pressure change data while walking. We acquire over ten steps gait data at once. The pressure change of the first step is unstable. Therefore, we employ eight steps (four right steps and four left steps) pressure change data from second step. Fig. 4 shows an example of the pressure change data of right and left foot.

III. FEATURE EXTRACTION For personal identification, we extract gait features from

each foot pressure change data. As shown in Fig. 4 pressure change of right foot and left foot are two different patterns. The pressure change of one step of the right foot is similar to the other steps of the right foot, and the pressure change of one step of the left foot is similar to the other steps of the left foot. We extract features, FX,S, from one step. Where, the notation X denotes an index of right foot (X = R) or left foot (X = L), and the notation s denotes an index of steps (s = 1, 2, 3 and 4). First, we extract one step pressure change data. Second, we extract peak pressure value, PAX,S, PBX,S, PCX,S and PDX,S and these times, TAX,S, TBX,S, TCX,S and TDX,S from pressure change of each circular tip. Third, we calculate intervals, I1X,S, I2X,S and I3X,S. They are calculated from TA X,S to TBX,S, from TAX,S to TCX,S, and from TAX,S to TDX,S, respectively. We employ these pressure values and times as gait features. Finally, we extract these eleven features {PAX,S, PBX,S, PCX,S, PDX,S, TAX,S, TBX,S, TCX,S, TDX,S, I1X,S, I2X,S and I3X,S} from all X and S. Fig. 5 shows example of one step pressure changes and features, FX,S of one

step.

IV. PERSONAL IDENTIFICATION We consider N registered persons and a walking person. We

identify the walking person using Euclidean distance. Preceding personal identification, we construct template, Sn,X, for each person. Where, notation n denotes registered person number (n = 1, 2, ... , N). For construct a template data, Sn,X, we calculate mean value of each features from learning pressure change data of target person n. We construct the template, Sn,X, for all registered person.

Next, we describe personal identification using Euclidean distance. First, we extract features FX,S from pressure change data of the walking person. Second, for all n, X and S, we calculate Euclidean distance dn,X,S between features FX,S and template Sn,X. For each step, we choose a registered person with the shortest dn,X,S as a candidate of the walking person. Then, we count the number of the steps. The number is called by the selected step number. Finally, we identify the walking person by the larger number. For example, if person A has 5 as the selected step number, and person B has 3 steps as the selected step number, then we identify the walking person as person A. If more than two persons have the largest number, identification of this walking person is failed. Moreover, if the

Figure 1. Experimental system.

Figure 2. Positions of sensing tips of pressure sensor.

Figure 3. Example of the fixed sensor shoes.

(a) Right foot pressure change data

(b) Left foot pressure change data

Figure 4. Example of acquired pressure change data.

arch

arch

PtA

PtB PtC

PtD

PtA

PtB PtC

PtD

Arch

Arch

Left foot

Right foot

121121

largest number less than threshold parameter Th identification of this walking person is failed. Table I shows an example of personal identification. When Th is set to 4, � is identified as person A, because A has the largest number of selected steps and over Th. � is not identified because the largest selected steps number is lower than Th. � is also did because B and C are the largest.

V. EXPERIMENTAL RESULTS In our experiment, we employed 10 volunteers in Table II.

We acquired pressure change data 10 times for the each. From these 10 pressure change data, we randomly chose 3 pressure data as learning data to construct template data. We use the other 7 pressure data for test data. Accuracy varies with choice of learning data. Therefore, this evaluation procedure was repeated 120 times in the style of cross validation method. For the performance assessment, we used false rejection rate (FRR) and false acceptance rate (FAR). FRR and FAR are most popular measures used in the biometric personal identification. FRR and FAR are defined by Eq. (1) and Eq. (2), respectively. FRR is concerned with the number of instance defined as an authorized individual being falsely rejected by an identification system. FAR is concerned with the number of instances defined as an unauthorized individual being falsely accepted by an identification system. The higher FRR decreases the

user-friendliness and higher FAR increases the risk of intrusion. 100 [%]Numbers of false rejectionsFRR

Numbers of authorized attempts� � (1)

100 [%]Numbers of false acceptancesFARNumbers of impostor attempts

� � (2)

In the experiment, we compared the proposed method that using 8 channels pressure sensor (Octsens) and a personal identification method that using sock liner type pressure sensor (Nitta Corp., F-Scan). The identification method [13] using F-Scan is similar to the proposed method; it acquires foot pressure distribution change. F-Scan is a foot shaped as Fig. 6 and this sensor size is 100mm × 285mm× 0.1mm (width× length × thickness). The sampling frequency of F-Scan is 50Hz and it obtains 1140 pixels in a foot. Fig. 7 shows an example of foot pressure distribution data. This method calculates pressure change data as total pressure change in each area. The identification method extracts same gait features and identifies walking person using Euclidean distance.

We identified volunteers by these identification methods. We performed these systems for the threshold (Th = 1, 2, 3, 4, 5, 6, 7 and 8). Table III and Fig. 8 shows identification result of each threshold. From this result, we can see that when threshold increases, FRR became high and FAR became low. In general, FRR and FAR need to be low. When Th less than 3, FRR and FAR of the proposed method using Octsens were low value and FRR was 0.83% and FAR was 0.02%. The proposed method obtained lower FRR and FAR than the identification method using F-Scan. The sampling frequency of Octsence (100Hz) is higher than that of F-Scan (50Hz). The sensor number (4 channels) of Octsence is much lower than that (X cannel) of

Figure 5. Example of one step pressure change and its features.

TABLE I. EXAMPLE OF PERSONAL IDENTIFICAITON

Walking person

Number of the selected steps A B C D E

� 7 0 0 0 1 � 0 3 2 2 1 � 0 4 4 0 0

TABLE II. VOLUNTEERS INFORMATION

Volunteer Gender Age [year]

Height [cm]

Weight [kg]

A Male 24 171 57 B Male 23 175 76 C Male 23 172 62 D Male 23 167 53 E Male 23 171 68 F Male 24 165 50 G Male 23 173 63 H Male 23 172 55 I Male 22 180 95 J Female 24 160 50

Figure 6. Sock liner type pressure sensor; F-Scan.

Figure 7. An example of foot pressure distribution data.

0 1Time [sec]

255

PA

PC

PD

PB

TA TB TC TD

I1

I2 I3

PtA

PtB

PtC

PtD

Pres

sure

val

ue [n

.u.]

122122

F-Scan. Considering these, our system was superior to the F-Scan system.

VI. CONCLUTION We proposed a biometric personal identification method

using 8 channels pressure sensor (Octsens). We employed both right and left foot pressure change while walking as the biometric behavioral feature. In this method, a wearable pressure sensor acquired the foot pressure change. Sensing tips of the sensor were fixed on optimal position for each volunteer. For personal identification, we extracted eleven gait features from each step pressure change data, and we calculate Euclidean distance between features of walking person and template. We identified person by the majority-decision using the Euclidean distance. In the experiments, we employed 90 volunteers. As the result, we identified with 0.83% in FRR and 0.02% in FAR. As the comparison results of our proposed method with an identification method using sock liner type pressure sensor (F-Scan), the proposed method obtained lower FRR and FAR. So, the proposed method succeeded in reducing sensor size and having high identification performance.

In the future, we will employ more volunteers, and check identification performance of our method. We will add other

identification methods or features, and identify person by multimodal processing.

ACKNOWLEDGMENT The authors would like to appreciate Mr. Takeshi Yamakawa

for his great supports.

REFERENCES [1] S. Liu and M. Silverman, “A practical guide to biometric security

technology,” IEEE IT Proc., 2001, pp. 27-32. [2] A. K. Jain, A. Ross, and S. Prabhakar, “An introduction to biometric

recognition,” IEEE Trans. on circuits and systems for video technology, Vol. 14, No. 1, pp. 4-20, 2004.

[3] Biometrics security consortium. “Biometric security technology handbook”, Biometrics security consortium, Ed., 1st Edn., Ohmhsa, Tokyo, Japan (2006). (in Japanese)

[4] A. Jain and S. Pankanti, “Biometrics systems: anatomy of performance”, IEICE Trans. Fundamentals, E-84-D, No. 7, pp. 788-799, 2001.

[5] J. W. Jung, Z. Bien, and T. Sato, “Person recognition method using sequential walking footprints via overlapped foot shape and center-of-pressure trajectory,” IEICE Trans. Fundamentals, Vol.E87-A, No. 6, pp. 1393-1399, 2004.

[6] J. W. Jung, D. J. Kim, and Z. Z. Bien, “Realization of personalized services for intelligent residential space based on user identification method using sequential walking footprints,” Journal of Sytemics, Cybernetics and Informatics, Vol. 8, No.6, pp. 90-95, 2010.

[7] G. Qian, J. Zhang, and A. Kidane, “People identidication using gait via floor pressure sensing and analysis,” Proc. of the 3rd European Conf. on Smart sensing and Context, pp. 83-98, 2008.

[8] L. Wang, H. Ning, T. Tan, and W. Hu, “Fusion of static and dynamic body biometrics for gait recognition,” IEEE Trans. on circuits and systems for video technology, Vol. 14, No. 2, pp. 1-11, 2004.

[9] D. K. Wagg and M. S. Nizon, “On automated model-based extraction and analysis of gait,” Proc. of the 6th IEEE Int. Conf. on Automatic Face and Gesture Recognition, pp. 11-16, 2004.

[10] J. Mantyjarvi, M. Lindholm, E. Vildjiounaite, S.-M. Makel, and H. A. Ailisto, “Identifying users of portable devices from gait pattern with accelerometers,” IEEE, ICASSP’05, Vol. 2, pp. ii/973-ii/976, 2005.

[11] T. Yamakawa, K. Taniguchi, T. Momen, S. Kobashi, K. Kondo, and Y. Hata, "Biometric personal identification based on foot pressure change," Int. Conf. on Soft Computing and Human Sciences, pp. 83-86, 2007.

[12] T. Takeda, K. Taniguchi, K. Asari, K. Kuramoto, S. Kobashi, and Y. Hata, "Biometric personal authentication by one step foot pressure distribution change by load distribution sensor," Proc. of 2009 IEEE Int. Conf. on Fuzzy Systems, pp. 906-910, 2009.

[13] T. Takeda, K. Taniguchi, K. Asari, K. Kuramoto, S. Kobashi, and Y. Hata, "Biometric Personal Authentication by One Step Foot Pressure Distribution Change by Fuzzy Artificial Immune System," Proc. of 2010 IEEE World Congress on Computational Intelligence, pp. 844-849, 2010.

[14] T. Takeda, K. Taniguchi, K. Asari, K. Kuramoto, S. Kobashi, and Y. Hata, "Biometric personal identification by dinamics of sole pressure at walking," Proc. of 2010 World Automation Cong., 2010.

[15] M. D. Addlesee, A. H. Jones, F. Livesey, and F. S. Samaria, “The ORL active floor,” IEEE Personal Communications, IEEE Wireless Communications, Vol. 4, Issue 5, pp. 35-41, 1997.

[16] R. J. Orr and G. D. Abowd, “The smart floor: a mechanism for natural user identificatoin and tracking,” Proc. of CHI 2000, 2000.

[17] T. Yamakawa, K. Taniguchi, K. Asari, S. Kobashi, and Y. Hata, "Biometric personal identification based on gait pattern using both feet pressure change," Proc. of 2008 World Automation Cong., 2008.

TABLE III. PERSONAL IDENTIFICATION RESULT

Th Proposed method (Using Octsens) Using F-Scan method

FRR [%] FAR [%] FRR [%] FAR [%] 1 0.83 0.02 1.61 0.182 0.83 0.02 1.61 0.183 0.83 0.02 1.90 0.184 0.86 0.02 3.13 0.075 2.07 0.00 6.93 0.026 9.45 0.00 15.46 0.007 27.33 0.00 47.67 0.008 60.19 0.00 56.20 0.00

Figure 8. FRR and FAR of the proposed method with Th change.

0.00

0.01

0.02

0.03

0

10

20

30

40

50

60

70

1 2 3 4 5 6 7 8

FRR[%]

FAR[%]

FRR

[%]

FAR

[%]

Threshold parameter, Th

123123

Fabrication and Characterization of Lateral-Directional Actuator Using Pb(Zr, Ti)O3

Thin FilmsK. Kanda, Member IEEE, Yu Akashi, Takashi Saito, Non-Member, IEEE, T.Fujita, K. Higuchi, and K.

Maenaka, Member, IEEE

�Abstract— Surface MEMS (microelectromechanical systems),

which is fabricated by the microprocessing techniques based on thin film deposition and processing, are incorporated into consumer electronics and automobiles. The structures of the surface MEMS having thin shape with small aspect ratio are difficult to drive to the lateral direction. The conventional electrostatic actuators have comb electrodes, resulting in a complex structure. This research addresses on the actuators available to move in a lateral direction by using thin-film Pb(Zr,Ti)O3. Test cantilevers with two segmented top electrodes on the piezoelectric film are fabricated and the driving characteristics are evaluated. Applications of electrical voltages with reversed phases to the segmented top electrodes at the resonant frequency with lateral oscillation mode moved the cantilever with large displacement of 7 �m. A non-linear effect in the frequency response was also observed for the large applied voltages.

Index Terms— Microactuators, Microelectromechanical systems, Piezoelectric films

I. INTRODUCTION

URFACE micromachining for CMOS technology has been applied to microelectromechanical systems (MEMS)

technology. The MEMS devices are called as surface MEMS in contrast to bulk micromachining. The advantage of the surface MEMS is good compatibility to CMOS circuits. Various microsensors are combined with the peripheral circuitry and integrated (e.g. microaccelerometers [1] and gyroscopes [2]).

On the other hands, the structures of the surface MEMS are quite thin because the surface micromachining is the combination of thin-film depositions and processing. Therefore surface MEMS typically have a small aspect ratio. The small aspect ratio compels actuators to move to a normal direction. For an actuator application such as resonator in gyroscopes, the multiplication of the degree of freedom (DOF) in the movement with miniaturization is a trend in the development. The electrostatic principle is conventionally used for the actuators. However, the comb electrodes in the electrostatic actuators complicate the shape of the device structure and generate troubles such as sticking of structures and failures due to particles. On the other hands, piezoelectric thin films have been

attracted attention for MEMS devices because of its simple actuation principle. Numerous devices using piezoelectric materials for MEMS devices have been reported [3]. We also have demonstrated output voltage operation on a mechanical structure by using series-connected elements of ferroelectric Pb[Zr,Ti]O3 (PZT) thin films for tri-axis accelerometer [4, 5] and sputtering technique of PZT/PZT bimorph structures [6]. Although PZT has excellent piezoelectric characteristics such as high electro-mechanical coupling factor and high piezoelectric d constant, the actuator applications are more preferable for PZT due to its high permittivity rather than sensors.

Manuscript received May 31, 2012. This research was partially supported by the Ministry of Education, Science, Sports and Culture, Grant-in-Aid for Young Scientists (B), 23760233, 2011.

K. Kanda is with University of Hyogo and Japan Science and Technology Agency, Himeji 6712280 Japan (phone: 81-79-267-4870; fax: 81-79-266-6225; e-mail: [email protected]).

Y. Akashi, was with University of Hyogo, Himeji, Hyogo 6712280 Japan (e-mail: [email protected])

As mentioned above, typical electrostatic actuators can have lateral- and normal- direction DOF dependent on the design. However, PZT actuators have generally the only DOF (i.e.normal (thickness-) directional actuation). Multiplication of DOF for the piezoelectric actuator can enhance the applicability of piezoelectric materials to devices. Actuation in lateral (width-) direction of a cantilever composed of piezoelectric thin film and microfabricated substrate is reported in this paper.

T. Saito is with Japan Science and Technology Agency, Himeji 6712280 Japan (e-mail: [email protected])

II. LATERAL ACTUATION OF CANTILEVERT. Fujita is with University of Hyogo and Japan Science and Technology Agency, Himeji 6712280 Japan (e-mail: [email protected]).

A. Operation Principle and Device Fabrication K. Higuchi is with Japan Science and Technology Agency, Himeji 6712280 Japan (e-mail: [email protected])

K. Maenaka is with University of Hyogo and Japan Science and Technology Agency, Himeji 6712280 Japan (e-mail: [email protected]).

Figure 1 is a schematic illustration of the actuation principle of piezoelectric actuator movable to lateral direction. A bottom

S


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.11

124


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.11

124

electrode, a piezoelectric thin film and width-directionally segmented two top electrodes are stacked on a cantilever. Electrical voltages are applied between top and bottom electrodes. The reversed voltages are applied to segment two electrodes each other. The piezoelectric thin film underneath the segment top electrodes strains dependent on the location, i.e. positive or negative applied voltage results in tensile or

compressive stress in the film. The fabrication process flow of the piezoelectric lateral

actuator is shown in Fig. 2. In this study, piezoelectric Pb(Zr, Ti)O3, PZT, was employed for the piezoelectric material because of its excellent piezoelectric characteristics. The deposition and processing techniques of PZT has been reported [7]. 4 inch SOI wafer with the thicknesses of device, buried oxide, and handle layers of 8, 1, and 300 �m, respectively, was thermally oxidized to make SiO2 on the substrate in a steam atmosphere at the temperature of 1000˚C. Pt/Ti thin film with the thickness of 100/10 nm was sputtered to form interfacial and bottom electrode layer at the temperature of 530˚C. PZT thin film with the thickness of 3 �m was sputtered at the temperature of 475˚C. Sequentially top Pt/Ti electrode was also sputtered at a room temperature. To fabricate cantilever structure, top Pt/Ti, PZT, and bottom Pt/Ti thin films were sequentially dry-etched. The etching process has been reported in [7]. Deep reactive ion etchings were employed to form the cantilever shape. The fabricated cantilever is shown in Fig. 3. Top electrodes are segmented into two area divided at the width center. The dimensions of the cantilever are 820 �m long, 100 �m wide and 8 �m thick in the cantilever (except for Pt/Ti/PZT).

Figure 1 Schematic illustration of the actuation principle for lateral-direction actuator

Figure 2 Fabrication process flow for lateral actuation cantilever

B. Finite Element Analysis For a static actuation of the cantilever, it is predicted that the

lateral displacement is quite small. To enhance the displacement, the cantilever should be vibrated at a resonant frequency. To estimate the static displacement of the cantilever, piezoelectric analysis by using finite element analysis software, IntelliSuite, was conducted. The displacement field when static voltages of 0.5 V and -0.5 V are individually applied to the PZT thin films is calculated. Fig. 4 is the lateral and normal displacements obtained. It was confirmed that the displacement was quite small.

To obtain resonant frequency at laterally moving modal shape, modal analysis was also conducted. Figure 5 indicates the resulting modal shape and the frequency. At third mode resonant oscillation, lateral motion was observed.

Figure 3 Fabricated cantilever device.

Displacement direction

(a) First modal shape (14.8 kHz) (b) Second modal shape (92.6 kHz)

(c) Third modal shape (142 kHz) Figure 5 Result of modal analysis.

-3.6

-0.25

0

0

Displacement direction Displacement direction

nm nm

(a) Lateral displacement contour (b) Normal displacement contour

Figure 4 Displacement profile calculated from finite element method.

125125

C. Frequency Respose To examine the oscillation characteristic of a fabricated

actuator, normal and lateral displacement was measured by using Doppler vibrometer when the reverse voltages with 1 V are applied to the top electrodes. The frequency response around third mode resonant frequency is indicated in Fig. 6. At the third resonant frequency of about 129.6 kHz, clear peak of the tip displacement was observed. The displacement in lateral direction was much larger than that in normal direction. The quality factor of the cantilever was measured to be 812.

D. Non-Linear Behavior The driving voltage was enlarged to obtain large

displacement of the actuator. Figure 7 indicates the influence of the driving voltage on the frequency response. The results demonstrate definite non-linear effect which the stiffness of the cantilever becomes small with increase in the displacement. Although the clear cause of the non-linear effect is unknown

without further investigation, torsional movement of the cantilever might play an important role. The lateral displacement and resonant frequency when the driving voltage is changed are plotted in Fig. 8. The non-linear effect was observed at the voltages over 0.5 V. The tip displacement during downward and upward frequency sweeping is demonstrated in Fig. 11. Such the instable frequency response is typical non-linear behavior. For downward frequency sweeping oscillation, the lateral tip displacement of 7 �m was obtained.

III. CONCLUSION

To actuate thin structure in lateral direction by using piezoelectric manner, PZT unimorph cantilever with segmented top electrodes are fabricated. Reverse voltage application to width-directionally segmented top electrodes at a resonant frequency of lateral oscillation mode actuate the cantilever in the lateral direction. When the lateral

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

129.8 129.9 130 130.1 130.2

Tip

disp

lace

men

t[�m

]

Driving frequency [kHz]

Out-of-plane

Width-direction

Figure 6 Comparison of the tip displacement in lateral and normal directions around third mode resonant frequency. The driving voltage is 1 V.

0

0.5

1

1.5

2

2.5

129.6 129.8 130 130.2 130.4

Driving frequency (kHz)

Late

ralt

ipdi

spla

cem

ent(

�m)

0

0.5

1

1.5

2

2.5

129.8

129.9

130

130.1

0 1 2 3 4

Late

ralt

ipdi

spla

cem

ent[

�m]

Res

onan

tfre

quen

cy [k

Hz]

Driving voltage [Vrms]Figure 8 Driving voltage vs. Lateral tip displacement and resonant frequency.

Driving voltage: 15 V

0

1

2

3

4

5

6

7

8

127.5 128 128.5 129 129.5 130

Driving frequency (kHz)

Late

ralt

ipdi

spla

cem

ent(

�m)

Figure 7 Relationship between lateral tip displacement and driving frequency for various driving voltages. Non-linear characteristics are observed.

Figure 9 Non-linear characteristics observed for upward and downward frequency sweeps

126126

displacement of the cantilever becomes large, non-linear frequency characteristics were observed. For a downward frequency sweep with the applied voltage of 15 V, the lateral tip displacement reached about 7 �m. It is experimentally verified that simple structure can generate a large lateral motion by using PZT.

REFERENCES

[1] C. Hierold, A. Hildebrandt, U. Naher, T. Scheiter, et al., “A pure CMOS surface micromachined integrated accelerometer,” in Proc. IEEE Micro ElectroMechanical Systems Workshop (MEMS’96), 1996, pp. 174–179.

[2] J. Bernstein et al. Feb. 1993. “A micromachined comb-drive tuning fork rate gyroscope,” in Proc. IEEE Micro ElectroMechanical Systems Workshop (MEMS ’93), 1993, pp. 143-148.

[3] S. Toroliear-McKinstry and P. Muralt, “Thin Film Piezoelectrics for MEMS,” J. Electroceram. Vol. 12, 2004, pp. 7-17.

[4] K. Kanda, Y. Iga, J. Matsuoka, T. Fujita, K. Higuchi, and K. Maenaka, “Multiplication of Sensor Output Voltage Using Series-Connected Piezoelectric Elements ,” IEEJ Trans. SM., Vol. 130, 2010, pp. 124-129 in Japanese.

[5] K. Kanda, Y. Iga, J. Matsuoka, Y. Jiang, T. Fujita, K. Higuchi, and K. Maenaka, “A Tri-Axial Accelerometer with Structure-Based Voltage Operation by Using Series-Connected Piezoelectric Elements,” ProcediaEngineering, Vol. 5, 2010, pp. 894-897.

[6] T. Saito, K. Kanda, Y. Iga, K. Higuchi, and K. Maenaka, “Intelligent Tri-Axis Accelerometer Capable of Output-Voltage Operation Upo Mechanical Structure,” Proc. the 28th Seensor Symposium, 2011, pp. 177-181 in Japanese.

[7] K. Kanda, Y. Iga, T. Hashimoto, T. Fujita, T. Higuchi, and K. Maenaka, “Microfabrication and Application of Series-Connected PZT Elements,” Procedia Chemistry, Vol. 1, 2009, pp. 808-810.

127127

Abstract—To encourage drivers to keep their safety driving, it’s

effective to analyze their in-vehicle video sequences objectively. In this paper, we propose an evaluation method of driving behavior by using computer vision techniques. Here we use road-markings to estimate road environment and vehicle speed. When the car approaches pedestrian crosswalk or stop sign, the driver should reduce the vehicle speed carefully. By several evaluating experiments, we verify and discuss effectiveness of our method. Keywords—Computer vision; Image processing; Object

detection; Pattern matching; Performance evaluation; Vehicle safety

I. INTRODUCTION Recently, an in-vehicle data camera which is an equipment

for recording video, audio, speed, etc. becomes common. For this reason, cases that its video sequences are used for evidence at law and insurance companies that establish discount system for cars installed in-vehicle data camera have been increasing gradually.

It is the greatest benefit of using in-vehicle data camera that we can clearly comprehend scenes of traffic accidents. Understanding what happened in the accident is terribly difficult for even the person who had it, because it happens in a moment. Hence, it’s not easy to elucidate its situation and cause on the basis of the verbal evidence of the driver and the passengers. Recording as we can comprehend objectively what happened in traffic accidents with in-vehicle cameras helps to reveal causes of traffic accidents and to make up quarrels. However, people who drive non-business-use cars rarely have traffic accidents, so there are not many opportunities to utilize videos of in-vehicle data camera except in traffic collisions. Thereupon, the evaluation method of driving behavior using videos of in-vehicle data camera has attracted attention.

Some companies which work business-use cars like buses, taxies, and trucks already use videos of non-accident for safe driving education of their employees, and actually the software which analyzes datum of in-vehicle data camera is already in use [1]. However, there’re some differences in the performance between cameras for business-use cars and ones for non-business-use cars. On the one hand to evaluate driving behavior in detail is easy because cameras for business-use cars can record a lot of datum, e.g. video, audio, speed, acceleration, location, turn signal, braking, time of idling, mileage, top speed,

but on the other hand it’s hard because ones for non- business-use cars record only minimum of data necessary, e.g. video, audio, speed and location with Global Positioning System (GPS), to sell cheaper.

Accordingly, the purpose of this study is developing a system to evaluate driving behavior with various data extracted from in-vehicle data cameras for non-business-use cars. This aims at providing drivers an indication which enables to review their driving behavior objectively, that is, need not proceed in real-time. Furthermore information we can get is only video (Fig. 1) and speed supplied by GPS (the sampling rate is 1 Hz).

Here we report an evaluation method using road-markings, and this paper consists of 6 sections. Section II presents pre-processing. In Section III, we describe how to detect road-markings from video of in-vehicle data camera, after that present the method of stop time estimation in Section IV. Section V is the experimental result of our method, and finally, in Section VI, this paper is drawn to a conclusion.

II. PRE-PROCESSING

A. Lens Distortion Correction Since almost all in-vehicle data camera use wide-view

fish-eye lens, captured images are distorted as shown in Fig. 2 (a). Thus we have to correct image distortion by camera calibration like Fig. 2 (b).

We use Zhang’s method [2] for camera calibration. It’s a process to calculate parameters, A and [R|t]. A is called the

An Evaluation Method of Driving Behavior by In-Vehicle Data Camera

Ryo ISHIZAKI, Masakazu MORIMOTO, and Kensaku FUJII, Division of Computer Engineering, University of Hyogo

Fig. 1. Captured image by in-vehicle data camera


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.39

293


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.39

293

camera intrinsic matrix which is influenced by properties of cameras such as focal length, Pixel Aspect Rate of Charged Coupled Device. [R|t], called the extrinsic parameters, is the rotation and translation which relates the world coordinate system to the camera coordinate system. After that, we can relate a 3D point, denoted by M, and a 2D point, denoted by m, with the parameters calculated; in short, we can correct the distortion. The relationship between M and m is given by

� � ,~|~ MtRAms � (1) where s is an arbitrary scale factor. However, this process causes losses of information at edges of images.

B. Perspective Transformation When we watch only road surface like road-markings,

captured images by in-vehicle data cameras have a lot of unnecessary information. Moreover, the scale variants in appearance among each flame of them are serious. Then, we calculate the matrix of a perspective transform from 4 pairs of the corresponding points, and transform front-view images to top-view images (Fig. 3). The upper side of the translated image is blurred because of the interpolation.

C. Template Matching Template matching is one of techniques in pattern matching

for finding small parts of an image which match a template image; image region where the degree of similarity is the highest is regarded as the result when compare a template image and an inspection image. If a pixel value of the point of the inspection image is I(W,H) and a pixel value of the point of the template image is T(w,h), the correlation value R(x,y) is defined by

� ��

� � � �.

',''',''

',''','',

',' ','22

','

�

��

yx yx

yx

yyxxIyxT

yyxxIyxTyxR (2)

This R(x,y) is a degree of similarity calculated by Normalized Cross-Correlation (NCC) [3], computes after brightness normalization. The value by NCC ranges from -1 to 1. As an inspection image is more similar to a template image, so R(x, y) becomes higher.

III. ROAD-MARKING DETECTION Road-markings can be found by executing template matching

in images applied distortion correction and perspective transform.

General road-markings, as shown in Fig. 4, are signs written on road surface by white or orange paint which notify us of traffic rules. Consequently, datum excepting road surface aren’t necessary, so we transform captured images to bird’s-eye-views.

(a) Original image

(b) Corrected image

Fig. 2. Lens distortion correction

(a) Front-view image (b) Top-view image

Fig. 3. Perspective transformation

Fig. 4. General road-markings

294294

Template matching cannot deal with a target object in an inspection image whose scale or angle difference from a template image. To solve this matter, in advance, we rotate and scale up or down template images automatically. Additionally, the top of translated images by perspective transformation gets blurred because of the interpolation. Therefore, in case of using template images of clear outlines, road-marking detection is difficult because the correlation value returned by template matching becomes lower. Then, we shade off template images moderately by Gaussian blur also known as Gaussian smoothing to enable the detection in blurred images. Although shape and scale variation arise in top-view images from jolting due to bumpy roads, this problem is also improved. An example of the template images created automatically is Fig. 5.

The degree of similarity given by template matching ranges from -1 to 1. And the more it gets close to 1, the more a template image resembles an inspection image. We judge the existence of road-markings by setting a threshold using it.

If a road-marking is found, we check what number the frame is from the first and speed-datum of GPS around the checked frame. Then, we can evaluate the driving behavior passing over road-markings with self-vehicle speed.

IV. STOP TIME ESTIMATION We can get the self-vehicle speed when the car passes over

road-markings by extracting the speed information from its GPS data. However, some in-vehicle data cameras have no GPS. And if they have GPS, the sampling rate of GPS is 1 Hz in most cases. This means that the GPS data isn’t enough to analyze driving behavior strictly. Then, as shown in Fig. 6, we estimate whether the car stopped or not and its stop time by partitioning images into 10x10 blocks and calculating their correlation coefficients between previous frames and next frames. Here, we except plain blocks, like road-surfaces (all black) and the sky (all blue or all gray), according to the variance of their pixel values in each block because all information of blocks is not always useful.

V. EXPERIMENT AND RESULT

A. In-vehicle Data Camera In the following experiment, we use DR-2000 (Fig. 7) as an

in-vehicle data camera. This camera has 1.5 megapixel and records VGA size (640x480) video frames at a rate of 30 fps. Moreover, it has GPS and acceleration sensor; records the speed and trajectory of itself for every second as GPS information. In the experiments, we use this GPS measured velocity as ground truth value.

B. Road-marking Detection Here, we experimented with 2 kinds of road-markings; one

warns existence of a crosswalk which has no pedestrian signals,

and another directs a temporary stop. Template images which aren’t blurred are shown in Fig. 8,

Fig. 9 shows the results. In contrast, Fig. 10 presents blurred template images, and the results are Fig. 11. In the inspection image, Region of Interest (ROI) was set a narrow area only the middle (1/3 of the width) to detect no road-marking on opposite lanes. The results in Fig. 9 and Fig. 11 verify that to use blurred template images improves the performance of the detection in the top of the inspection images. We can also know the effect of Gaussian blur by Fig. 12 that shows the relationship between the Gaussian kernel size and the detection rate. The figure says the range of the kernel sizes which have the highest detection rate is from 11 to 17. And if we make the kernel size higher, it defeats its own purpose. In addition, we observe that this method is effective in the rain and nighttime (Fig. 13).

Fig. 5. Template images created automatically

Fig. 6. Image partitioned into 10x10 blocks

Fig. 7. In-vehicle data camera (DR-2000)

295295

(a) Crosswalk-notice (b) Temporary stop

Fig. 8. Non-blurred template image

(a) Crosswalk-notice (b) Temporary stop Fig. 9. Result of detection by non-blurred template image


Fig. 10. Blurred template image


Fig. 11. Result of detection by blurred template image

Fig. 12. Relation between Gaussian filter kernel size and direction rate

Fig.13. Result of detection in the rain and nighttime

Fig. 14. Result of stop time estimation

Fig. 15. Result of stop time estimation (GPS couldn’t record 0km/h)

C. Stop Time Estimate The result of stop time estimation when the car passed over

road-markings which indicate a temporary stop is shown in Fig. 14. In this figure, Correlation Coefficient means the median of correlation coefficients when invalid correlation coefficients were removed by using the variance of pixel values in each block, and the car was stopping while it’s nearly 1. As a result of confirming stop time by visual observation, we validated that we could know accurate stop time than GPS by this method. Additionally, we also validated that it’s possible to estimate stop time if GPS could not record 0km/h because of its shortness as shown in Fig. 15.

D. Evaluation by Road-marking and Self-vehicle Speed We confirmed validity of the evaluation method using

road-markings by analyzing driving behavior when the car passed over the road-marking which indicates a temporary stop.

First, we executed road-marking detection on the sequences captured by the in-vehicle data camera and recorded what number the last detected frame is from the first (detected-frame’s number). Next, we examined the correspondence between each frame and the self-vehicle speed by extracting the information of time and speed from GPS data. Finally, we confirmed the self-vehicle speed the car passing over the temporary stop indicator by examined the relation between detected-frame’s number and the extracted datum. In relation to datum of 2 drivers, Fig. 16 is images of

296296

detected-frame’s number, and Fig. 17 shows images when the speed of both datum became 0km/h (after 6 seconds from the detected-frame’s number). Their velocity change after 10 seconds from the detected-frame’s number, if the time of detected-frame’s number is 0s, is illustrated in Fig. 18. Fig. 18 shows us that the driver A accelerated and decelerated more careful than the driver B. Otherwise, B is regarded as checked the safety in the neighborhood warily because the stop time of B is longer than A. On the other hand, the result of stop time estimation about the same video is shown in Fig. 19. Here, tA is A’s stop time and tB is B’s stop time. This figure shows, in actuality, A’s stop time is longer unlike the result of Fig. 18 which is evaluated from only GPS data.

These results concluded that evaluation by this method is a valid approach.

VI. CONCLUSION In this study, we did road-marking detection and stop time

estimation of self-vehicle. And we proposed an evaluation method using road-markings and self-vehicle speed for an effective utilization of videos captured in-vehicle data camera except in traffic accidents. We implemented road-marking detection by template matching with top-view images and stop time estimation by using correlation coefficients which is better than GPS. We confirmed validity of the evaluation method which is based on road-marking detection, stop time evaluation and speed information included GPS data.

In future works, we have to establish some evaluation methods by combine our method with other methods like people detection, new indicators to evaluation driving behavior such as the distance between cars, and an evaluation system of driving behavior by putting them together as a whole.

REFERENCES [1] Ministry of Land, Infrastructure, Transport and Tourism, “Business

Report about Utilizing Model of In-Vehicle Camera in 2008,” Mar. 2009, pp. 8-9.

[2] Zhengyou Zhang, “A Flexible New Technique for Camera Calibration,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(11):1330-1334, 2000.

[3] J. P. Lewis, “Fast Normalized Cross-Correlation,” 1995.

(a) A (b) B Fig. 16. Images of detected-frame’s number

(a) A (b) B

Fig. 17. Images when the speed of both datum became 0km

Fig. 18. Velocity change during 10 seconds after detected

Fig. 19. A’s and B’s results of stop time estimation

297297

�Abstract— Asthma causes the bronchus inflammation, and

makes breathing impossible. In worst case, asthma causes death bydyspnea. If we can predict cause asthmatic attacks, they can prevent from asthmatic attacks. Therefore, asthmatic attacks prediction system is desired. As a prediction system using timeseries data, there is Fuzzy-AR model that can consider multi factors. In this paper, we propose a prediction method of the number of asthmatic attacks on next month based on Fuzzy-ARmodel. The proposed method considers weather factors; temperature, atmospheric pressure and humidity data. Thismethod is applied to asthmatic attacks data from Himeji city Medical Association. As a comparison method, AR model is applied to same data. The experimental results shown that the proposed method predicts the number of asthmatic attacks better than AR model.

Index Terms—asthmatic attack, autoregressive model, healthcare system, time-series data.

I. INTRODUCTION

HE number of asthmatic attacks in Japan was over 1 million in 2005 [1]. Asthma causes the bronchus inflammation, and

makes breathing impossible. In worst case, asthma leads death by dyspnea.

The asthmatic attacks have relations with temperature, atmospheric pressure, humidity, ticks, air pollution and so on. Therefore the number of the asthmatic attacks fluctuates in ayear. Fig. 1 shows the total number of asthmatic attacks in Himeji city from 2001 to 2010. According to Fig. 1, the number of asthmatic attacks fluctuates in a year [2]. There is a difference in the number of the asthmatic attacks each generation. Fig. 2 shows the number of asthmatic attacks by generation in Himeji city from 2001 to 2010. According to Fig. 2, the number of asthmatic attacks from 1 to 4 years old is greatest number. And over 65 years old is the second [2]. It is observed difference in each generation. If we can predict asthmatic attacks increases in

Manuscript received May 31, 2012 Y. Kaku is with Graduate School of Engineering, University of Hyogo,

Hyogo, Japan, and Himeji Initiative in Computational Medical and Health Technology, Hyogo, Japan (corresponding author to provide phone: +81-79-267-4986; e-mail: [email protected]).

K. Kuramoto, Syoji Kobashi and Y. Hata are with Graduate School of Engineering, University of Hyogo, Hyogo, WPI Immunology Frontier Research Center, Osaka University, Osaka, and Himeji Initiative in Computational Medical and Health Technology, Hyogo, Japan.

each generation, asthmatics can prevent from asthmatic attacks.Therefore asthmatic attack prediction system is desired.

As a prediction system using time series data, there is autoregressive (AR) model [3]-[5]. For example, AR model predict stock prices, body weight [6] and so on. AR model predict from only past data. On the other hand, there is Fuzzy-AR model that can consider other factors [7]-[10].

In this paper, we propose a prediction method of the number of asthmatic attacks on next month based on Fuzzy-AR model. The proposed method considers weather factors; temperature, atmospheric pressure and humidity data. The method is applied to asthmatic attacks data from Himeji city Medical Association [2]. Temperature, atmospheric pressure and humidity data are acquired from Japan Meteorological Agency [11]. The proposed method and AR model method is applied to asthmatic data, and compare the prediction result.

II. PREDICTION METHOD BY AR MODEL

AR model predicts by using time series data. AR model is defined by (1).

)()()()(1

tuitxiatyp

i��

�

(1)

Here, y(t) denotes a predicted value of time t, x(t) does a true value, a(i) does AR parameter, u(t) does white noise, and p does the order. In this study, a(i) is calculated by using Yule-Walker equation defined by (2).

��

�

�

�

�

��

�

�

�

��

�

�

�

��

��

)(

)2()1(

)(

)2()1(

)0()2()1(

)2()0()1()1()1()0(

pR

RR

pa

aa

RpRpR

pRRRpRRR

��

�

��

�

�

(2)

Here, the notation R(i) denotes auto covariance function. Moreover, the order p is determined on the basis of Akaike’s information criterion (AIC) [12], [13]. In this study, the order pis determined from the time-series data of the number of learning month. The method calculates the p which minimizes the value of AIC. AIC is calculated by (3).

kLAIC 2)ln(2 �� (3)Here, L denotes maximum likelihood and k does number of free parameter. Maximum likelihood L is changed by p.

Predict Time Series Data for The Number of Asthmatic Attacks in Himeji by Fuzzy-AR Model

Yusho Kaku, Student Member, IEEE, Kei Kuramoto, Member, IEEE, Syoji Kobashi, Member, IEEEYutaka Hata, Fellow, IEEE

T


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.31

314


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.31

314

III. PROPOSED METHOD BASED ON FUZZY-AR MODEL

Fuzzy-AR model can consider multi factors. In this study, temperature, atmospheric pressure and humidity are considered to predict the number of asthmatic attacks. It adds fuzzy rule to the prediction method by AR model which is explained in section II.

The proposed method predicts number of asthmatic attacks based on Fuzzy-AR model. Fig. 3 shows the flowchart of this method. First, the method calculates AR parameter a(i) from number of asthmatic attacks by using (2). Second, the method calculates degree of weather μ from weather data. Finally, the method predicts the number of asthmatic attacks next month using a(i) and μ.

The method considers three knowledge of weather data as shown in the following knowledge. Knowledge 1: when temperature falls, asthmatic attacks are

caused. Knowledge 2: when atmospheric pressure falls, asthmatic

attacks are caused. Knowledge 3: when humidity rises, asthmatic attacks are

caused. According to these knowledge, the following fuzzy IF-THENrules are derived. Rule 1: IF temperature of the month to predict is lower than

mean temperature of the same month from weather data base,

THEN the degree of the temperature μT is high. Rule 2: IF the atmospheric pressure of the month to predict is lower than mean atmospheric pressure of the same month from weather data base,THEN the degree of the atmospheric pressure μP is high. Rule 3: IF humidity of the month to predict is higher than mean humidity of the same month from weather data base,THEN the degree of the humidity μH is high.

According to rule 1 to 3, fuzzy membership functions are defined show in Fig. 4. In Fig. 4, the notice T denotes temperature of prediction month, Tave does mean temperature of weather data base, P does atmospheric pressure of prediction month, Pave does mean atmospheric of weather data base, Hdoes humidity of prediction month, and Have does mean humidity of weather data base.

μ is calculated from μT, μP and μH by (4).

3HPT ��

� (4)

According to past data, the number of asthmatic attacks tends to converge to the mean. Therefore we make uniform random number as correction term r(t) to close to mean value. To make r(t) rule is defined by (5).

��

�otherwise

gnnniftstr meanlastmean

0*)(

)( (5)

Here, s(t) denotes uniform random number from 0 to value that is subtracted nlast from nmean. nmean does mean asthmatic attacks data base of prediction month. nlast does the number of asthmatic attacks of the year that is last year of the month to predict. gdoes rate of fluctuation . In this study, g is 20%.

Using (1), (4) and (5), Fuzzy-AR model is defined by (6).

0

400

800

1200

1600

2000

2001 2003 2005 2007 2009

asth

mat

ic a

ttack

s(ca

ses)

yearsFigure1. The number of asthmatic attacks from 2001 to 2010

0

5000

10000

15000

20000

25000

30000

0 1-4 5-9 10-14 15-19 20-24 25-44 45-64 65-

asth

mat

ic a

ttack

s(ca

ses)

age(year)Figure2. Total of the number of asthmatic attacks from 2001 to 2010

Figure3. Flowchart of the proposed method.

End

Start

Calculate AR parameter

Number ofasthmatic attacks

Calculate fuzzy degree of weather

Weather data

The number of asthmatic attacks

of next month

315315

)(*)5.0()()()()(1

trwtuitxiatyp

i��

�

(6)

Here, w denotes a weighting value, and r(t) does correction term. In our study, w is 150.


The method predicted the number of asthmatic attacks next month from 1 to 4 years old and over 65 years old respectively.In our study, every past month data of the asthmatic attacks was used to predict the number of asthmatic attacks of next month. In order to predict the number of next month of 2006 to 2010, the proposing method uses the number of asthmatic attacks and weather data from 2001 to 2005 as database.

Fig. 5 shows the prediction results the number of asthmatic attacks from 1 to 4 year with Fuzzy-AR model and AR model. According to Fig. 5, Fuzzy-AR model and AR model was able to predict the number of asthmatic attacks roughly.

Fig. 6 shows the prediction results the number of asthmatic attacks over 65 years old with Fuzzy-AR model and AR model. According to Fig. 6, Fuzzy-AR model and AR model was not able to predict the number of asthmatic attacks.

Table I tables the comparison of correlation coefficient andmean error results between ground truth value and the results from 1 to 4 years old of Fuzzy-AR model and AR model.

According to Table I, Fuzzy-AR model obtained 0.857 ± 0.037mean correlation coefficient and 45.7 (cases) mean error, and AR model obtained 0.800 ± 0.045 mean correlation coefficient and 46.5 (cases) mean error.

Table II tables the comparison of correlation coefficient and mean error results between ground truth value and the results over 65 years old of Fuzzy-AR model and AR model. According to Table II, Fuzzy-AR model obtained 0.191 ± 0.052 mean correlation coefficient and 46.1 (cases) mean error, and AR model obtained 0.118 ± 0.050 mean correlation coefficient and 45.3 (cases) mean error.

V. CONSIDERATION

Fig. 7 shows histogram of the order p from 1 to 4 years old.According to Fig. 7, frequency of the order 12 is highest. Therefore we are able to know that asthmatic attacks from 1 to 4 years old have 1 year cycle.

Fig. 8 shows histogram of the order p over 65 years old. According to Fig.8, frequency of order are scattered around 1 to 3. Therefore we are able to know that asthmatic attacks over 65 years old have monthly volatility.

Comparing prediction results from 1 to 4 years old and over 65 years old. The proposed method predicted the number of asthmatic attacks from 1 to 4 years old. However, it was not able to predict the number of asthmatic attacks over 65 years old. From these results, AR model is applicable to only data that have a cycle.

degree

1.0

0.5

0 T(K)TaveTave-CT Tave+CT

μT

Tx

STx(T) TEMPERATURE

(a) Temperature

degree

1.0

0.5

0 P(hPa)PavePave-CP Pave+CP

μP

Px

SPx(P) PRESSURE

(b) Atmospheric pressure

degree

1.0

0.5

0 H(%)HaveHave-CH Have+CH

μH

Hx

SHx(H) HUMIDITY

(c) HumidityFigure4. Fuzzy membership functions.

0

100

200

300

400

500

600

2006 2007 2008 2009 2010

asth

mat

ic a

ttack

s(ca

ses)

(year)

Ground truth valueFuzzy-AR modelAR model

Figure5. Prediction results from 1 to 4 years old with Fuzzy-AR model and AR model.

0

50

100

150

200

250

300

350

2006 2007 2008 2009 2010

asth

mat

ic a

ttack

s(ca

ses)

(year)

Ground truth valueFuzzy-AR modelAR model

Figure6. Prediction results over 65 years old with the proposed method and AR model.

316316

VI. CONCLUSIONS

In this study, we predicted the number of asthmatic attacks from 1 to 4 years old by using Fuzzy-AR model. This method considered temperature, atmospheric pressure and humidity data. As the experimental result, the correlation coefficient of Fuzzy-AR model between the predicted value and ground truth value was 0.857. And that of AR model was 0.800. By comparison with these results, the proposed method accurately predicted the number of asthmatic attacks from 1 to 4 years old than AR model. The proposed method was able to do a highly accurate prediction by adding weather factors.

In the future, we will add other factors to Fuzzy-AR model, because the proposed method was considered only weather. For example addition factors, there are location data, air pollution and so on. And we will try to predict the number of asthmatic attacks over 65 years old by other prediction model.

REFERENCES

[1] G Ministry of Health home page http://www.mhlw.go.jp/toukei/saikin/hw/kanja/05/05.html

[2] Himeji city Medical Association home page http://www.himeji-med.or.jp/iinkai/kougai/geppou.html

[3] J. Wang and T. Zhang, “Degradation prediction method by use of autoregressive algorithm,” IEEEICIT, pp. 1-6, April 2008.

[4] F. M. Bennett, D. J. Christini, H. Ahmed, K. Lutchen, J. M. Hausdorff and N. Oriol, “Time series modeling of heart rate dynamics,” Computers in Cardiology, p.273, September 1993.

[5] W.Gersch and T. Brotherton, “AR model prediction of time series with trends and seasonalities: A contrast with Box-Jenkins modeling,” Decision and Control including the Symposium on Adaptive Processes, vol. 19, p.988, December 1980.

[6] H. Tanii, H. Nakajima, N. Tsuchiya, K. Kuramoto, S. Kobashi and Y. Hata, “A Fuzzy Logic Approach to Predict Human Body Weight Based on AR model” Fuzz-IEEE, pp. 1022-1025, June 2011.

[7] R. Palaniappan, P. Reveendran, S. Nishida and N. Saiwaki, “Evolutionaryfuzzy ARTMAP for autoregressive model order selection andclassification of EEG signals,” Systems, Man and Cybernetics, vol. 5, p.3682, October 2000.

[8] J. Liu, J. Mo, S. Pourbabak, “Human cardiovascular system identification and application using a hybrid method of auto-regression and neuro-fuzzy inference systems,” Machine Learning and Cybernetics, vol 7, p.4107, August 2004.

[9] B. S. Chen, S. C. Peng and K. C. Wang, “Traffic Modeling, Prediction, and Congestion Control for High-Speed Networks: A Fuzzy AR Approach,” IEEE Transaction on Fuzzy Systems, vol. 8, pp. 491-508, October 2000.

[10] N. Watanabe, “A Fuzzy Rule Based Time Series Model,” Fuzzy Information, vol. 2, p.936, June 2004.

[11] Japan Meteorological Agency home page [12] http://www.jma.go.jp/jma/index.html H. Akaike, “A new look at the

statistical model identification,” IEEE Transactions, vol. 19, pp. 716-723, December 1974.

[13] R. Shibata, “Selection of the order of an autoregressive model by Akaike’s information criterion,” Biometrika, vol 63, pp. 117-126, June 1975

TABLE I. PREDICTION RESULTS FROM 1 TO 4 YEARS OLD WITH THE FUZZY-AR MODEL AND AR MODEL.

Fuzzy-AR model AR model

year correlation coefficient

meanerror(cases)

correlation coefficient

meanerror(cases)

2006 0.883±0.026 39.7 0.844±0.036 43.0

2007 0.891±0.028 39.6 0.834±0.032 57.1

2008 0.858±0.039 57.6 0.899±0.027 58.7

2009 0.782±0.049 61.8 0.791±0.045 39.1

2010 0.869±0.041 30.0 0.634±0.083 34.6

total 0.857±0.037 45.7 0.800±0.045 46.5

TABLE II. PREDICTION RESULTS OVER 65 YEARS OLD WITH THE FUZZY-AR MODEL AND AR MODEL.

Fuzzy-AR model AR model

year correlation coefficient

meanerror(cases)

correlation coefficient

meanerror(cases)

2006 0.193±0.053 52.3 0.170±0.047 50.4

2007 0.648±0.044 25.7 0.348±0.054 33.5

2008 0.040±0.049 59.7 0.142±0.046 53.5

2009 -0.165±0.075 40.3 -0.181±0.060 41.1

2010 0.239±0.041 52.4 0.112±0.044 47.8

total 0.191±0.052 46.1 0.118±0.050 45.3

0

10

20

30

40

50

60

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17

Freq

uenc

y

The order p (month)

0

10

20

30

40

50

60

1 2 3 4 5 6 7 8 9 10 11 12

Freq

uenc

y

The order p (month)Figure7. Histogram of the order p from 1 to 4 years old Figure8. Histogram of the order p over 65 years old

317317

Abstract— The heating effect by microwave irradiation for the

ecologic system was investigated by using cubic phantoms and plants. For the accurate evaluation, the irradiation was performed in the hollow waveguide. The relation between the microwave absorption and surface temperature increase by the microwave heating were theoretically estimated by the principle of heat transfer. These results were also confirmed by the experiment performed with 2.45GHz microwave irradiation. The absorption coefficient of the cubic phantom with sides 30mm long in the 109.2mm×54.6mm hollow waveguide was measured to be about 0.38 at 2.45GHz. The surface temperature increase of the cubic phantom by the irradiation of 2W microwave was around 18 degree, which is agree with the calculation in error of about 20%. Some kinds of vegetables were also used in the irradiation experiment and observed the surface temperature increase by the heating effect.

Index Terms—Microwave irradiation, Ecologic material, Heat transfer, Hollow Waveguide, Phantom

I. INTRODUCTION

n recent years, the applications of wireless transmission of the electric power using electromagnetic waves have been

extensively studied. For example, the Solar Power Satellite system requires the microwave power transmission from the satellite to the ground.[1] It is necessary for the practical use of such systems to inquire the ecology of microwaves and the influence on the ecological system.

The evaluation of microwave irradiation effects on the plants has been performed by exposing to microwaves in space.[2] Influence of microwave power for ecology system is however difficult to be quantitatively analyzed by such method because the microwave absorption in the plants cannot be exactly estimated owing to the microwave scattering in all directions. The heating by the microwave irradiation is the most important effect as the ecology of microwave and is directly related to the microwave absorption of the ecologic materials.

Here, we employ a microwave irradiation method where ecological materials are placed into a hollow waveguide and the

Manuscript received May 14, 2012. Xu Jingfan, T. Kawai, A. Enokihara are with Department of Electrical

Engineering and Computer Sciences, Graduate school of engineering, University of Hyogo, 2167 Shosha, Himeji, Hyogo 671-2280, Japan (+81-79-267-4872; [email protected]).

H. Murakami is with Energy Technology Research Institute, National Institute of Advanced Industrial Science and Technology (AIST), 1-1-1 Umezono, Tsukuba, Ibaraki, 305-8568 Japan.

microwaves scattered by the object can be quantitatively determined by the scattering parameters measured by the waveguide network. Thereby the absorption coefficient and the heat rate in the object can be exactly estimated.

In this paper, the heating effect by microwave irradiation in the hollow waveguide was considered by using cubic phantoms and plants. The relation between the microwave absorption and surface temperature increase by the microwave heating were theoretically estimated by the principle of heat transfer. The surface temperature increases of the cubic phantom and some kinds of plants by the irradiation of 2W microwave are also shown.

II. HEATING BY THE MICROWAVE IRRADIATION

A. Microwave absorption coefficients of the cubic phantoms

The microwave absorption coefficients of the cubic phantoms were measured by using the waveguide network at 2.45GHz. The phantom consists of saline water and powered polyethylene to emulate electric properties of the living body. The permittivity and the loss tangent were measured to be 25 and 1.0 respectively. The gelatinous phantom material was filled up in a cubic container made of thin plastic sheets. Fig. 2 shows the photograph of the cubic phantoms. These cubic phantoms, with sides 20mm, 30mm and 40mm long, were used for the experiment. Each phantom was fixed at the center of the hollow waveguide WRJ-2, of which cross section is 109.2mm×54.6mm. The outline of the experimental setup is shown in Fig. 1. Numerical calculation of the microwave absorption coefficient was also performed using electromagnetic simulator (HFSS).

Network Analyzer

a

Phantom

Waveguide WRJ-2

Microwave

Coaxial Waveguide Converter Fig. 1. Experimental setup for the measurement of

absorption cofficient

Investigation of Heating Effect by Microwave Irradiation for Phantoms and Plants

Jingfan Xu, Tadashi Kawai, Member, IEEE, Akira Enokihara, Member, IEEE, and Hiroshi Murakami

I


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.32

318


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.32

318

a=40mma=30mm

a=20mm

Fig. 2. The photograph of the cubic phantoms

The absorption coefficients A at 2.45GHz are shown in table

1. These coefficients are determined using the S11 and S21 scattering parameters detected by the network analyzer. It is seen that the A increases with the side length of the cube and that this feature was observed both in the experiment and in the numerical calculation. The measured absorption coefficient of the cubic phantom with sides 30mm long was about 0.38.

Table 1. Absorption coefficients

Length a of the side Calculated value Experimental value

[mm] of A of A

20 0.197 0.182

30 0.416 0.378

40 0.643 0.64

B. Heat transfer from the cubic body to air

The heat transfer rate q from the cubic body with sides a long is presented by where , T and T0 are heat transfer coefficient, surface temperature of the body and temperature of air, respectively. Here we assume that all faces of the cube are in the same temperature. The heat transfer coefficient can be determined using the principle of heat convection [3]. Fig. 3 shows the calculation result of the heat transfer rates q as functions of T, where T0 is 20 degree centigrade. Note that the q increases with the side length a of the cube and that this feature similarly exist in the relation between absorption coefficient A with a in Table 1.

0

2

4

6

0 20 40 60 80 100 120

Hea

t tra

nsfe

r rat

eq

[W]

Surface temperature Tω [deg.]

a=20mma=30mma=40mm

Fig. 3. Calculation results of heat transfer rate of cubic body

C. Surface temperature increase by microwave irradiation power

When the cubic body is irradiated by microwaves, the cubic body will be heated up by the microwave absorption. The temperature of the body will become a constant value at the state of the steady heat flow where the heat rate and the heat transfer rate are in the same value. The amount of the temperature increase at the steady state depends on both the heat rate in body and the heat transfer coefficient from body to air.

The heat rate qm by the microwave absorption is obviously presented as qm =AP, where P is a irradiated microwave power and q=qm at the steady state. Fig. 4 shows the surface temperature T as functions of P at T0 of 20 degree centigrade. It is seen that the irradiation power dependence of T does not almost depend on side length a of cube. This is attributable to the fact that the relation between q and a is similar to that between A and a as mentioned before. The surface temperature increase ΔT=TT0 is estimated to about 23 degree at 2W irradiation power.

0

20

40

60

0 0.5 1 1.5 2 2.5

Surf

ace

tem

pera

ture

Tω

[deg

.]

Input power P [W]

a=20mma=30mma=40mm

Fig. 4. Surface temperature as functions of input power

III. MICROWAVE IRRADIATION EXPERIMENT

Fig. 5 shows experimental setup of microwave irradiation for cubic phantoms and some kinds of plants. The irradiated object is placed at the center of the hollow waveguide WRJ-2. The input microwave power was fixed at 2W.

Power Meter

Coaxial Waveguide Converter

phantom

Waveguide WRJ-2

Isolator Amplifier

MicrowaveOscillators 2.45GHz

Attenuator

Fig. 5. Experimental setup of microwave irradiation

Fig. 6 shows the irradiation time dependences of the surface

temperature increase ΔT for the three cubic phantoms. The surface temperatures of all faces of each cube were measured at every ten minutes using an infrared ray radiation thermometer. The average value of temperatures of all surfaces of each cube

20ω )( aTTq

319319

is used as the T. The surface temperature increases ΔT at the steady state are shown in Table 2. ΔT by the microwave irradiation is around 18 degrees. The measured values are about 20 % lower than those of calculation for all three cases. This may be attributable to the assumption in the calculation that the all six faces of the cubic phantoms are in the same temperature.

0

5

10

15

20

0 25 50 75 100

Surf

ace

tem

pera

ture

in

crea

se Δ

T[de

g.]

Time [min]

a=20mma=30mma=40mm

Fig. 6. Surface temperature increase of cubic phantom

Table 2. Surface temperature increase ΔT [deg.] of phantom

Length a of the side of the cube

[mm]Numerical calculation Measurement

20 22.41 18.330 23.14 18.640 23.69 17.8

Next we attempted to apply the same experimental setup to the microwave irradiations for some kinds of plant body. Table 3 shows the surface temperature increase, absorption coefficient and mass density.

Table 3. Experimental results for plants

Materials

Surface temperature increase ΔT

[deg.]

Absorption coefficients

A

Mass density[g/cm3]

Wood(Cube a=30mm) 4.4 0.0103 0.521

Tomato(Sphere d=30mm) 12.2 0.117 0.687

Cactus(Sphere d=30mm) 13.7 0.239 0.645

Potato(Cube a=30mm) 10.5 0.370 1.117

Carrot(Cube a=20mm) 8.0 0.109 1.071

Carrot(Cube a=30mm) 9.2 0.347 1.176

It is seen that the absorption coefficient of each plant related with the mass density for three cases of cubic body with sides 30mm long. This should means that the mass density depends on the water content and that the included water must

dominantly cause the microwave absorption. In cases of potato and carrot, the evaporation of water from their surfaces might suppress the temperature increase. The numerical calculation of the surface temperature increase ΔT is difficult for these plants since the ΔT depends on many factors other than the absorption coefficient, size, shape, and surface conditions. The microwave absorption coefficient however is an effective factor to evaluate the microwave irradiation effect for plants.

IV. CONCLUSION

The heating effect by microwave irradiation at 2.45 GHz for the ecologic materials was investigated. For the quantitative consideration, the irradiation was performed in the hollow waveguide.

REFERENCES [1] H.Matsumoto, “Research on solar power satellites and microwave power

transmission in Japan,” IEEE Microwave magazine, 3, 36-45, 2002. [2] H.Murakami, K.Komiyama, Y.Kato and I.Kudo, ”Recent Progress in

Long-Duration Microwave Exposure Facility,” 52nd International Astronautical Congress, IAF-01-R.1.08, 2001.W.-K. Chen, Linear Networks and Systems (Book style). Belmont, CA: Wadsworth, 1993, pp. 123–135.

[3] W.H. McAdams, Heat Transmission. New York: McGrow-Hill, 1954.

320320

Hyper-

thermia

WLAN

WLAN

WLAN

WLAN

Cellarphone

Disturbance Collision

Conventional wall or partition

Disturbance

(a)

Hyper-

thermia

WLAN

WLAN

WLAN

WLAN

Cellarphone

Disturbance Collision

Disturbance

Frequency selective shielding wall

(b)

Fig. 1. (a)The environment of electro-magnetic wave inn a conventional

hospital, (b) in a new hospital using the frequency selective shielding wall.

Abstract— In a hospital, many equipments use electro-magnetic

wave. The equipment which uses for energy transmits very high

power. This high power electro-magnetic wave disturbs other

electronic equipments which use the same frequency. To avoid the

disturbance, frequency selective shielding technologies are

effective. The laminated structure consisting of metal wire grid

and dielectric layer are studied. The characteristics of these

structures are controlled by the parameters such as wire diameter

and spacing. The characteristics are calculated by the use of the

transmission line model, and measured. All the measured results

show good match with the calculated results.

Index Terms—Electro-magnetic Wave, Metal Wire Grid,

Frequency Selective Shielding, Transmission.

I. INTRODUCTION

N 19TH

century, the concept of electro-magnetic wave was

predicted by J. C. Maxwell and proved experimentally by H.

R. Hertz. The applications of the electro-magnetic wave are

commonly used in our life. The applications of the

electro-magnetic wave are categorized by communication,

measurement, and energy utilization. Wireless communication

represented by a cellar-phone is the most notable application in

the first category; radar and microwave oven are in the second

and the third category respectively. In a hospital, all of the three

categories of application are utilized. A cardiac telemetry is

commonly used, and this is the typical application in the first

category. MRI scanner is the typical one in the second category.

And an electrical scalpel and a hyper-thermia are the typical

ones in the third category.

MRI is settled in a shielded room, but other equipments such

as an electrical scalpel and a hyper-thermia are commonly used

in non-shielded room, and the electro-magnetic wave from the

equipments in the third category is intensive and there is the risk

that it gives interference to other medical equipment. For

example, when a hyper-thermia works in a non-shielded room,

the electro-magnetic wave disturbs the communication of

wireless LAN system which uses the same frequency band.

Shielding is one important technique to avoid this disturbance.

On the one hand, wireless LAN systems are detrimental to

each other. When the traffic of the wireless LAN becomes large,

T. Iwai is with Kawasaki Techno-Reserch, Inc., Osaka,

Japan(corresponding author to provide ian-iwai@ maia.eonet.ne.jp).

S. Yamamoto, S. Aikawa, and K. Hatakeyama are with Graduate School of

Engineering, University of Hyogo, Himeji, Japan.

the collision of the demand of communication often occurs and

the through rate becomes small. To avoid this collision,

shielding technologies are useful, but shielding using metallic

plate reflects the electro-magnetic wave of all frequency. To

satisfy our demand of using cellar phones or TV in that room,

the frequency selective characteristic is preferable. For example,

a frequency selective shielding wall which shows low pass

characteristic of cut off frequency 2GHz is preferable.

In this paper, we present a frequency selective structure using

metal wire grids and dielectric layers.

Electro-magnetic Shielding Technologies in

Hospital

Tohru Iwai, Shinichiro Yamamoto, Satoru Aikawa and Kenichi Hatakeyama

I


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.33

321


978-0-7695-4884-5/12 $26.00 © 2012 IEEE

DOI 10.1109/ICETET.2012.33

321

Z0 Z0 Z0 Z0 ZwZw

b/2 b/2 d/2 d/2

Ywg

Fig.3. Equivalent circuit model of matal wire array sandwiched by thin

dielectric films.

(a) (b)

Fig. 2. Matal wire array (a), metal wire grid (b), and the structure sandwiched

by 2 thin dielectric films.

II. ELECTRO-MAGNETIC WAVE CHARACTERISTICS OF

LAMINATED STRUCTURE MADE OF METAL WIRE GRID AND

DIELECTRIC SHEET

A. Metal wire grid

The characteristics of metal wire grid which placed metal

wire to the equal distance are discussed when the

electro-magnetic wave is irradiated perpendicularly. Fig.2

shows a metal wire grid and a metal wire array. If the direction

of the electric field is parallel to the wires of metal wire array,

the characteristics of metal wire array is equal to that of metal

wire grid. In this study, we use metal wire array. The diameter

and spacing of wire are d and a, respectively. To maintain the

array structure metal wire array is sandwiched by 2 thin

dielectric films which thickness is b/2.

Using transmission model, this wire array is represented as an

inductance which is inserted in parallel, and the equivalent

circuit model is represented in fig.3. Ywg is represented by

equation(1).

⋅−==

d

aaZ

j

ZY

g

wg

π

λ

ln

1

0

0 (1)

Using this transmission model, the effective relative

permittivity εr and effective relative permeability µr are

calculated.

( )

( )XZ

dbjZ

dbZjY

m

mwg

r

r 2

0

0

0

2

2

=∆+∆

∆+∆+=

ε

ε

µε (2)

( )( ){ } YdbYjZ

dbj mwgrr =∆+∆+⋅

+−= − µ

πλ

µε 0

10 1cosh)2

(3)

Using equations(2) and (3), εr and µr are represented by

equation(4).

XYZr 0=ε , X

Yr =µ (4)

Fig.4 shows the calculated the effective relative permittivity

εr . This figure shows that the effective relative permittivity εr

becomes negative.

B. Laminated structure

The laminated structure consisting of metal wire grid and

dielectric layer are studied. Fig.4 shows the laminated structures.

When the thickness of the dielectric sheet is small shown in

Fig. 4. Calculated effective relative permittivity εr .

Wire grid layer Dielectric sheet layer

(a) (b) (c)

Fig.４. Laminated structures consisting of metal wire grid and dielectric sheet,

(a) metal wire grid is placed on one side, (b)metal wire grid is placed between

two dielectric sheets, (c) metal wire grids are placed on both side of dielectric

sheet.

322322

TABLE 1

PARAMETERS OF SIMULATION.

spacing： a a1:2.5 mm

a2:3.5 mm

a3:7.0 mm

diameter：d 0.2 mm

dielectric film thickness：ｂ/2 0.1 mm

relative permittivity：εw 3.2

dielectric sheet thickness：t 1.76 mm

relative permittivity：εd 6.8

4 6 8 10 12 14 16 18−20

−10

0

Frequency (GHz)

|Γ|,|Τ

| (d

B) |Γ|

|Τ|

a1

a2

a3a1

a2

a3

(a)

4 6 8 10 12 14 16 18- 20

- 10

0

|Τ|

|Γ|

a1

a2

a3

a3a2

a1

Frequency (GHz)

|Γ|，

|Τ| (d

B)

(b)

Fig. 5. Reflection and transmission coefficients of the laminated structures

consisting of metal wire grid and dielectric sheet, (a) calculated results, (b)

measured results.

fig.4(a), at the frequency where the mean value of permittivity

of the laminated structure becomes equal to 1, the reflection

becomes minimum and the transmission becomes maximum.

But the thickness becomes large, the reflection becomes large

and the transmission becomes small.

C. Metal wire grid is placed between two dielectric sheets.

The structure shown in Fig.4(b) is studied and the parameters

are shown in Table 1. The dielectric films are made of PET and

the dielectric sheets are made of glass.

Fig.5 shows the reflection and transmission coefficients of

this structure, (a) shows the calculated results, and (b) shows the

measured results. In the case of spacing of a3, the reflection is

nearly 0dB and the transmission is very low at the low frequency,

but, at 6GHz the reflection becomes -∞dB small and the

transmission becomes 0dB. At nearly 10GHz, the reflection

becomes large and the transmission becomes small. At 14GHz,

the reflection becomes small and the transmission becomes 0dB

again. At the spacing 3.5mm, the frequency interval between 2

frequencies where the reflection becomes -∞dB, and the wide

frequency band where the transmission is around 0dB exists. At

the spacing 7.0mm, the transmission does not become 0dB and

the minimum value of reflection becomes large.

Fig.5(b) shows the measured results of this laminated

structure. All the results show very good match with the

calculated results.

This laminated structure gives very wide frequency band

where the transmission is around 0dB.

D. Metal wire grids are placed on both side of dielectric

sheet.

The structure shown in Fig.4(c) is studied and the parameters

are shown in Table 2. The calculated transmission coefficients

are shown in fig.6. A characteristic to completely match at two

frequencies in 3-18GHz is provided. The tendency that the

transmission band of the low frequency side is narrower is

recognized. When spacing grows large, the matching frequency

decreased, and the tendency that the the transmission band

becomes wider is recognized. As frequency becomes low, the

transmission coefficient approaches -∞dB. The transmission

5 10 15- 30

- 20

- 10

0

Frequency (GHz)

Tra

nsm

issio

n (

dB

)

2.5mm

3.5mm

5.0mm

7.0mm

Fig. 6. Frequency characteristics of the transmission coefficients at the

spacings of 2.5, 3.5, 5.0, 7.0 mm.

TABLE 2

PARAMETERS OF SIMULATION.

diameter：d 0.2 mm

dielectric film thickness：ｂ/2 0.1 mm

relative permittivity：εw 3.2

dielectric sheet thickness：t 5.8 mm

relative permittivity：εd 6.8

323323

characteristics between both matching frequencies changes by

the spacing.

The measured transmission coefficients are shown in fig.7.

The spacing is 3.5mm. The calculated value of this figure is the

result including the dielectric loss. The characteristic depends

on a dielectric loss at a peak that the transmission is not 0dB and

is not for mismatching. And the error of the matching frequency

depends upon the gap between the metal wire grid and the

dielectric layer because the specimen is held horizontally during

the measurement.

III. CONCLUSION

In a hospital, many equipments use electro-magnetic wave.

The equipment which uses for energy transmits very high power.

This high power electro-magnetic wave disturbs other

electronic equipments which use the same frequency. To avoid

the disturbance, frequency selective shielding technologies are

effective. As an example of this frequency selective shielding

technologies, the laminated structure consisting of metal wire

grid and dielectric layer are studied.

The structure that a metal wire grid is sandwiched by tow

dielectric sheets shows the very wide frequency band where the

transmission is nearly 0dB, and the structure that a dielectric

sheet is sandwiched by tow metal wire grid shows the very sharp

transmission band. These characteristics are controlled by the

parameters such as wire diameter and spacing.

Other many structures show frequency selective shielding

characteristics. These technologies will provide solution of the

electro-magnetic environment where many electro-magnetic

waves are used for many purposes.

REFERENCES

[1] K. Hatakeyama, T. Tsutaoka, T. Kanemoto, S. Yamamoto, T. Iwai,

“Reflection and Transmission Characteristics of EM-Wave with the

Wire-Grid and Its Use as a Back Layer of EM-Wave Absorber,” IEICE

Trans., vol.J93-B, no.1, pp.101-111, (2010-1)(in Japanese)

[2] T. Iwai, T. Kanemoto, S. Yamamoto, K. Hatakeyama, “Reflection and

transmission characteristics of the multilayer structure consisting of the

metal grid and the lossy material layer,” IEICE Technical Report , vol.

110, no. 402, EMCJ2010-111, pp. 67-72 (2011-1)(in Japanese)

[3] T. Iwai, S. Yamamoto, K. Hatakeyama, “Reflection and Transmission

Characteristics of Multilayered Structure Consisting of Metal Wire Grid

and Dielectric layers,” IEICE Trans., Vol.J95-B No.3 pp.488-492 ,

(2012-3)(in Japanese)

[4] T. Iwai, S. Yamamoto, K. Hatakeyama, “EM Wave Transmission

Characteristics of Laminated Structure Consisting of Metal Wire

Grid-Dielectric-Metal Wire Grid,” IEEJ Trans. FM, Vol. 132, No. 5,

pp.385-386, (2012-5) (in Japanese)

5 10 15−30

−20

−10

0

Frequency (GHz)

Tra

nsm

issio

n (

dB

)

Measured

calculated

Fig. 7. Calculated and measured transmission coefficients at the spacing

3.5mm.

324324

医用生体情報解析グループ

畑豊教授（センター長）

藤井健作教授

佐藤邦弘教授

小橋昌司准教授

森本雅和助教

荒木望助教

Journal Papers

1. N. Yagi, Y. Oshiro, T. Ishikawa and Y. Hata, "YURAGI synthesis for ultrasonic human brain imaging," Journal of

Advanced Computational Intelligence and Intelligent Informatics, vol. 17, no. 1, pp. 74-82, 2013.

2. N. Yagi, Y. Hata and N. Shibanuma, "Fuzzy support system for total hip arthroplasty stem by ultrasonic

intraoperative measurement," Journal of Hindawi - Advances in Fuzzy Systems, 2012 (online).

3. Y. Hata, O. Ishikawa and H. Nakajima, "Elderly Health Care System of Systems by Non- Contacted Multi-Sensors,"

Journal of Public Health Frontier, online

4. S. Hayashi, T. Nishiyama, T. Fujishiro, N. Kanzaki, N. Shibanuma, S. Kobashi and M. Kurosaka, "Automated pelvic

anatomical coordinate system is reproducible for determination of anterior pelvic plane," Computer Methods in

Biomechanics and Biomedical Engineering, online first, 2012.

5. K. Sato, “Numerical reconstruction of high-resolution microscopic images for lens-less holographic microscopy”,

International Journal of Intelligent Computing in Medical Sciences and Image Processing, Vol. 4, No. 1, pp.

77-86 ,2012

6. K. Fujii, K. Kashihara, I. Wakabayashi, M. Muneyasu and M. Morimoto, “Application of simultaneous equations

method to ANC system with non-minimum phase secondary path”, IEICE Trans. Fundamentals, vol. E95, no. 7, pp.

1109-1116, 2012.

7. K. Fujii, Y. Okamoto, M. Itou, M. Muneyasu and M. Morimoto, “Two-channel adaptive algorithm for active noise

control system using simultaneous equations method”, Acost. Sci. & Tech, vol. 33, no. 6, pp. 339-347, 2012.

8. Y. Kuwahara, Y. Iwamatsu, K. Fujii, M. Muneyasu and M. Morimoto, “Normalization method of gradient vector in

frequency domain steepest descent type adaptive algorithm”, IEICE Trans. Fundamentals, vol. E95, no. 11, pp.

2041-2043, 2012.

9. 吉岡拓人，中野裕文，藤井健作，棟安実治, 森本雅和, “因果律に基づく音声強調２マイクロホンアレーシステム”, 日本音響

学会誌，vol. 63, no. 2, pp. 63-73, 2012.

10. Shintaro Nakatani, Yuhei Nishida, Nozomu Araki, Yasuo Konishi and Kunihiko Mabuchi; “A BCI Rehabilitation System

for Leg Paralysis Patients –Development and Examination of Wheelchair Shape Rehabilitation System–”, ICIC

Express Letters, Vol.6, No.11, 2809-2814, 2012.

Refereed Proceedings/口頭発表

1. A. Hashioka, K. Kuramoto, S. Kobashi, Y. Wakata, K. Ando, R. Ishikura and Y. Hata, "Fuzzy object shape model for

newborn brain MR image segmentation," Proc. of the 6th International Conference on Soft Computing and

Intelligent Systems and 13th International Symposium on Advanced Intelligent Systems, pp. 1253-1258, 2012.

2. T. Yagi, T. Takeda, K. Sueyoshi, Y. Oshiro and Y. Hata, "Fuzzy estimation system on gait independence level by

footprint dynamics," Proc. of the 6th International Conference on Soft Computing and Intelligent Systems and 13th

International Symposium on Advanced Intelligent Systems, pp. 1259-1264, 2012.

3. H. Tanii, H. Nakajima, N. Tsuchiya, K. Kuramoto, S. Kobashi and Y. Hata, "A fuzzy time-series prediction model with

multi-biological data for health management," Proc. of the 6th International Conference on Soft Computing and

Intelligent Systems and 13th International Symposium on Advanced Intelligent Systems, pp. 1265-1268, 2012.

4. N. Yagi, Y. Oshiro, T. Ishikawa and Y. Hata, "Human brain ultrasound-mediated diagnosis in emergency medicine

and home health care," Proc. of the 6th International Conference on Soft Computing and Intelligent Systems and

13th International Symposium on Advanced Intelligent Systems, pp. 1269-1274, 2012.

5. Y. Takashima, T. Ishikawa, S. Kobashi, K. Kuramoto and Y. Hata, "Ultrasonic evaluation of seminiferous tubules by

frequency map," Proc. of 2012 Fifth Int. Conf. on Emerging Trends in Engineering and Technology, pp. 7-12, 2012.

6. T. Takeda, S. Kobashi, K. Kuramoto and Y. Hata, "Biometrics personal identification by wearable pressure sensor,"

Proc. of 2012 Fifth Int. Conf. on Emerging Trends in Engineering and Technology, pp. 120-123, 2012.

7. S. Kobashi and J. Udupa, "Fuzzy Object Model Based Fuzzy Connectedness Image Segmentation for Newborn Brain

MR Images," Proc. of 2012 IEEE Int. Conf. on Systems, Man and Cybernetics, pp. 1422-1427, 2012.

8. T. Fujimoto, H. Nakajima, N. Tsuchiya, H. Marukawa, Y. Hata, "Human activity clasification by ECG and

accelerometers Aided by Fuzzy logic," Proc. of 2012 IEEE Int. Conf. on Systems, Man and Cybernetics, pp.

2036-2041, 2012.

9. M. Kuki, H. Nakajima, N. Tsuchiya, Y. Hata, "Human movement trajectory recording for home alone bu thermoplile

array sensor," Proc. of 2012 IEEE Int. Conf. on Systems, Man and Cybernetics, pp. 2042-2047, 2012.

10. K. Tsukuda, T. Egawa, K. Taniguchi and Y. Hata, "Average difference imaging and its application to ultrasonic

nondestructive evaluation of wind turbine blade," Proc. of 2012 IEEE Int. Conf. on Systems, Man and Cybernetics,

pp. 2601-2604, 2012.

11. H. Nakajima, T. Shiga and Y. Hata, "Systems health care - the aspect of home and medical care," Proc. of 2012 IEEE

Int. Conf. on Systems, Man and Cybernetics, pp. 2616-2621, 2012.

12. M. Endo, H. Nakajima and Y. Hata, "Simplified Factory Energy Management System Based on Operational Condition

Estimation by Sensor Data," Proc. of 8th IEEE Int. Conf. on Automation Science and Engineering, 2012.

13. H. Nakajima, T. Shiga and Y. Hata, "Systems health care --Coevolutionary integration of smart devices and smart

services," Proc. of 2012 Service Research and Innovation Insititute Global Conference, pp. 231-236, 2012.

14. Y. Hata and H. Nakajima, "A health care service to disaster survivors," Proc. of 2012 Service Research and

Innovation Insititute Global Conference, pp. 637-641, 2012.

15. H. Mita, S. Kobashi, K. Nagasawa, K. Nishiyama, H. Maeno, K. Kuramoto and Y. Hata, "A fusion method for marine

rader images using EM algorithm," Proc. of 2012 World Automation Cong., 2012. (online)

16. A. Hashioka, S. Kobashi, K. Kuramoto, Y. Wakata, K. Ando, R. Ishikura, T. Ishikawa, S. Hirano and Y. Hata, "Shape

and appearance knowledge based brain segmentation for neonatal MR images," Proc. of 2012 World Automation

Cong., 2012. (online)

17. N. Yagi, Y. Oshiro, T. Ishikawa and Y. Hata, "Ultrasonic image synthesis in fourier transform," Proc. of 2012 World

Automation Cong., 2012. (online)

18. Y. Hata, K. Kuramoto, S. Kobashi and H. Nakajima "A survey of fuzzy logic in medical and health technology," Proc.

of 2012 World Automation Cong., 2012. (online)

19. H. Tanii, H. Nakajima, N. Tsuchiya, K. Kuramoto, S. Kobashi and Y. Hata, "A Fuzzy-AR model to predict human body

weights," Proc. of 2012 IEEE World Congress on Computational Intelligence, pp. 2027-2032, 2012.

20. Y. Takashima, T. Ishikawa, K. Kuramoto, S. Kobashi and Y. Hata, "Ultrasonic Thickness Evaluation of Seminiferous

Tubule by Fuzzy Inference," Proc. of 2012 IEEE World Congress on Computational Intelligence, pp. 2033-2037,

2012.

21. Y. Hata, M. Nakamura, N. Yagi and T. Ishikawa, "Trans-skull ultrasonic doppler system aided by fuzzy logic," Proc. of

SPIE Defence, Security and Sensing 2012, pp. 840118-1-8, 2012.

22. H. Nakajima, N. Tsuchiya, T. Shiga and Y. Hata, "Systems health care ? daily measurement and lifestyle change,"

Proc. of SPIE Defence, Security and Sensing 2012, pp. 84011A-1-11, 2012.

23. Y. Hata, S. Kanazawa, M. Endo, N. Tsuchiya and H. Nakajima, "Health care system for heart rate and human

movement trajectory in house and its application to detect mental diseases," Proc. of SPIE Defence, Security and

Sensing 2012, pp. 840119-1-11, 2012.

24. K. Sato and Y. Iwayama, “Instantaneous shape measurement of moving surface with high accuracy by one-shot

digital holography”, Proc. of SPIE, Vol. 8281, pp. 0K1-0K6, 2012

25. K. Furuichi and K. Sato, “Real-time recording and reconstruction of moving 3D images using electronic holography”,

Proc. of SPIE, Vol. 8281, pp. 0L1-0L6, 2012

26. M. Otani and K. Sato, “Lensless holographic microscopy with high-resolving power for 4D measurement of

micro-organism swimming in water”, M. Otani, K. Sato, Proc. of SPIE, Vol. 8227, pp. 2901-2906, 2012

27. H. Kaneko, M.Morimoto, K.Fujii, “Vehicle Speed Estimation by In-Vehicle Camera”, Proc. of World Automation

Congress, 2012

28. T. Murai, M. Morimoto and K. Fujii, “A Visual Inspection System for Medical Capsules”, Proc. of IEEE International

Conference on Systems, man, and Cybernetics, pp.1433-1437, 2012.

29. R. Ishizaki, M. Morimoto and K. Fujii, “An Evaluation Method of Driving Behavior by In-Vehicle Data Camera”, Proc.

of Fifth International Conference on Emerging Trends in Engineering & Technology, 2012.

30. T. Murai, M. Morimoto and K. Fujii, “A Visual Inspection System for Prescription Drugs”, Proc. of Fifth International

Conference on Emerging Trends in Engineering & Technology, 2012.

31. K. Fujii, T. Yoshioka, Y. Kuwahara, M. Muneyasu and M. Morimoto, “A method steering a null toward a point in high

reverberate non-minimum phase acoustic space”, Proceedings of Acoustic, pp. 2833-2838, 2012.

32. 谷井秀謙，畑豊, "健康管理システムの開発～福島県立医科大学との共同研究の取り組み～," 兵庫県立大学研究テーマ集セ

レクション 2012，p.60, 2012.

33. 中嶋宏，志賀利一，畑豊, "System Health Care - デバイスとサービスの共通化的統合について," 第 35 回多値論理フ

ォーラム, 2012.

34. 畑豊, 八木直美, "ファジィ論理を用いた超音波医用画像システムの開発," 第 64 回多値論理研究会, 2012.

35. 佃光樹, 江川正人, 谷口和彦, 倉本圭, 小橋昌司, 畑豊, "超音波を用いた風力発電用ブレードの人工破損特徴検出,"

情報通信学会総合大会２０１２講演論文集, A-11-3, 2012.

36. 藤本達弘, 土屋直樹, 中嶋宏, 倉本圭, 小橋昌司, 畑豊, "運動強度と周波数に基づいた決定木によるヒト動作分類,"

情報通信学会総合大会２０１２講演論文集, D-7-2, 2012.

37. 九鬼正人, 中嶋宏, 土屋直樹, 倉本圭, 小橋昌司・畑豊, "天井設置型サーモパイルアレイによる歩行の動線抽出," 情報

通信学会総合大会２０１２講演論文集, D-7-3, 2012.

38. 小橋昌司, 喜多村祐里, 倉本圭, 下野九里子，谷池雅子, 畑豊, "ソフトコンピューティング画像処理による MR 画像を用

いたてんかん原性域検出," 映像情報メディア学会技術報告, Vol. 36, No. 54, pp. 61-64, 2012.

39. 橋岡亜弥, 小橋昌司, 倉本圭, 若田ゆき, 安藤久美子, 石藏礼一, 石川智基, 廣田省三, 畑豊, "新生児頭部 MR 画像から

の脳領域法におけるファジィ物体モデルの検討," 映像情報メディア学会技術報告, Vol. 36, No. 54, pp. 65-68, 2012.

40. 橋岡亜弥，小橋昌司，倉本圭，若田ゆき，安藤久美子，石藏礼一，石川智基，廣田省三，畑豊， "ファジィ物体形状モデ

ルを用いた新生児頭部 MR 画像からの脳領域抽出法," 第 28 回ファジィシステムシンポジウム講演論文集, pp. 767-772,

2012.

41. 小橋昌司, "医療用解析技術の現状と動向," 2012 医療用エレクトロニクス技術大全, pp. 59-62, 2012.

42. 三田北斗, "最尤推定法を用いたマルチ船舶レーダ統合システム," パーティクルフィルタ研究会, 2012.

43. 桑原佑輔，吉岡拓人，藤井健作，棟安実治，森本雅和, “非最小位相成分を含む音響空間の１点に向けて死角を構成する手法

に関する検討”, 電子情報通信学会応用音響研究会技術報告, EA2011-104. 2012.

44. 岸本涼鷹，吉岡拓人，藤井健作，棟安実治，森本雅和, “強残響下での音声強調が可能な２マイクロホンアレーシステムの提

案”, 電子情報通信学会応用音響研究会技術報告, EA2011-105, 2012.

45. 酒井龍矢，岩松祐輔，藤井健作，棟安実治*，森本雅和, “能動騒音制御装置における帰還系の同定法に関する検討”, 電子情

報通信学会応用音響研究会技術報告, EA2011-118, 2012.

46. 澤田拓也，桑原佑輔，藤井健作，棟安実治*，森本雅和, “巡回型と比巡回型縦続接続適応フィルタの音響エコーキャンセラへ

の適用に関する検討”, 電子情報通信学会応用音響研究会技術報告, EA2012-22, 2012.

47. 岸本涼鷹，桑原佑輔，吉岡拓人，藤井健作，棟安実治，森本雅和, “非最小位相音響空間においても音声強調が可能となる２

マイクロホンアレーシステムの提案”, 電子情報通信学会応用音響研究会技術報告, EA2012-83 , 2012.

48. 横谷祥史，酒井龍矢，藤井健作，棟安実治，森本雅和, “周波数領域マルチチャネルシステム同定アルゴリズムの提案”, 電子

情報通信学会応用音響研究会技術報告, EA2012-84 , 2012.

49. 岸田裕士，藤井健作，棟安実治，森本雅和, “非最小位相音響空間における残響低減法に関する検討”, 電子情報通信学会応用

音響研究会技術報告, EA2012-85 , 2012.

50. 藤井健作，岸本涼鷹，吉岡拓人，棟安実治，森本雅和, “再帰非再帰縦続接続適応フィルタのマイクロホンアレーへの適用”, 日

本音響学会 2012 年春季研究発表会,3-1-2 , 2012.

51. 藤井健作，桑原佑輔，棟安実治，森本雅和, “非最小/最小位相成分の分離法に関する検討”, 2012 年電子情報通信学会総合大

会, A-4-31 , 2012.

52. 藤井健作，横谷祥史，酒井龍矢，棟安実治，森本雅和, “周波数領域マルチチャネルシステム同定アルゴリズムの提案”, 2012

年電子情報通信学会ソサイエティ大会, A-4-8 , 2012.

53. 藤井健作，岸本涼鷹，桑原佑輔，吉岡拓人，棟安実治，森本雅和, “非最小位相音響空間における騒音低減マイクロホンアレ

ーに関する検討”, 2012 日本音響学会秋季研究発表会, 2-9-6 , 2012.

54. 前田大輔，森本雅和，藤井健作，”距離センサを用いた物体の分離・識別に関する検討”, 第 11 回情報科学技術フォーラム,

H-017, 2012.

55. 石崎遼，森本雅和，藤井健作，”車載カメラによる運転評価手法に関する検討”, 第 11 回情報科学技術フォーラム, H-022,

2012.

56. 村井貴昭，森本雅和，藤井健作，”処方薬剤の画像識別に関する研究”，第 11 回情報科学技術フォーラム, H-030, 2012.

57. 金子弘樹，森本雅和，藤井健作，”車載カメラを用いた車両速度推定に関する検討”, 平成 24 年度電気関係学会関西連合大会，

P-27, 2012.

58. 前田大輔，森本雅和，藤井健作，”距離センサを用いた物体の分離・識別に関する検討”, 平成 24 年度電気関係学会関西連合大

会，9amT-24, 2012.

59. 石崎遼, 森本雅和, 藤井健作, ”車載カメラによる運転評価手法の検討”, 平成 24 年度電気関係学会関西連合大会，9amT-25,

2012.

60. 村井貴昭，森本雅和，藤井健作，”薬剤の自動識別に関する研究”, 平成 24 年度電気関係学会関西連合大会，9amT-27, 2012.

61. 前田大輔，森本雅和，藤井健作，”距離センサを用いた物体の分離・識別に関する検討”, 動的画像処理実利用化ワークショッ

プ，IS3-12, 2012.

62. N. Araki, K. Inaya, Y. Konishi and K. Mabuchi; “An Artificial Finger Robot Motion Control Based on Finger Joint Angle

Estimation from EMG Singals for a Robot Prosthetic Hand System”, Proc. of the 2012 International Conference on

Advanced Mechatronic Systems, Tokyo, Japan, pp. 109-111, 2012.

63. 満渕邦彦, 中谷真太朗, 深山理, 荒木望; “BMI を応用した運動麻痺患者に対するリハビリテーション装置の開発 –その

意図と予備的実験結果について−”, 日本機械学会ロボティクス・メカトロニクス講演会 2012, 1A1-N11, 2012.

64. 稲谷健司, 荒木望, 小西康夫, 満渕邦彦; “生体信号に基づいた手指関節角度推定型ロボット義手の開発”, 日本機械学会ロ

ボティクス・メカトロニクス講演会 2012, 2P1-U11, 2012.

65. 中谷真太朗, 荒木望, 満渕邦彦; “BCI を用いた下肢麻痺患者用リハビリシステムの開発 –足漕ぎ車椅子型リハビリシステ

ムに関する基礎的検討−”, 第 51 回日本生体医工学会抄録集, p. 250, 2012.

66. 中谷真太朗, 新納弘崇, 横田将尭, 荒木望, 稲谷健司, 國本雅也, 鈴木隆文, 深山理, 石川正俊, 石垣博行, 下条誠, 満

渕邦彦; “マイクロスティミュレーション法を用いた筋電義手に対する感覚機能の付与”, 第 25 回日本マイクロニューログラ

フィ学会抄録集, p. 3, 2012.

67. 横田将尭, 中谷真太朗, 國本雅也, 荒木望, 満渕邦彦; “ヒトの筋紡錘からの求心性線維の発火パターンの指運動時におけ

る記録とその解析”, 第 25 回日本マイクロニューログラフィ学会抄録集, p. 5, 2012.

産業財産権

1. 佐藤邦弘、「ホログラフィック断層顕微鏡、ホログラフィック断層画像生成方法、およびホログラフィック断層画像用のデー

タ取得方法」、特願 2012-223690、平成 24 年 10 月 5 日

生体マイクロセンシンググループ

前中一介教授

藤田孝之准教授

神田健介助教

才木常正主任研究員（兵庫県立工業技術センター）

Journal Papers

1. Y. Jiang, K. Kanda, Y. Iga, T. Fujita, K. Higuchi, K. Maenaka, "Monolithic integration of Pb(Zr,Ti)O3 thin film based

resonators using a complete dry microfabrication process", Microsystem Technologies," s00542, published online

2. K. Kanda, J. Inoue, T. Saito, T. Fujita, K. Higuchi, K. Maenaka, "Fabrication and Characterization of Double Layer

Pb(Zr,Ti)O3 Thin Films for Micro-Electromechanical Systems", Japanese Journal of Applied Physics,"Vol. 51,

09LD12"

3. K. Kanda, T. Saito, Y. Iga, K. Higuchi, K. Maenaka, "Influence of Parasitic Capacitance on Output Voltage for

Series-Connected Thin-Film Piezoelectric Devices," Sensors, Vol. 12, pp. 16673-16684

4. X. Hao, Y. G. Jiang, H. Takao, K. Maenaka, K. Higuchi, "An Annular Mechanical Temperature Compensation Structure

for Gas-Sealed Capacitive Pressure Sensor," Sensors, Vol. 12, pp. 8026-8038

5. T. Namazu, J. Nakamura, K. Higuchi, K. Maenaka, "Novel Size Effect Phenomena in Single Crystal Silicon Nanowire,"

Journal of the Japanese Society for Experimental Mechanics, Vol. 12, No. 1, pp. 8-12

6. 村井康二, 保毛津林太郎, 林祐司, 才木常正, 樋口行平, 藤田孝之, 前中一介, "小型加速度センサによる水先人行動モニタリ

ングに関する基礎研究,” 神戸大学海事科学研究科紀要, 第 9 号 P1-6

7. 瀧澤由佳子，樋口行平，前中一介，才木常正, "歯装着型加速度センサを用いた音声認識の提案," 電気学会論文誌 C, Vol. 132,

No. 7, pp. 1206-1207

8. 濱田浩幸，前中一介, "生体張付材料としての絆創膏型ポリマーとその応用," Material Stage, Vol. 12, No. 2, pp.12-15


1. 宮川宜和，横松得滋，竹内悠，神田健介，藤田孝之，樋口行平，前中一介, "ハニカム SOI ウエハを用いた加速度センサ," 電

気学会交通・電気鉄道／フィジカルセンサ合同研究会,TER12-005/PHS12-005

2. 牧村賢一，橋本泰知，宮本亜聖，笠井一夫，神田健介，藤田孝之，前中一介, "MEMS リングレーザジャイロのためのミラー

作製," 電気学会交通・電気鉄道／フィジカルセンサ合同研究会,TER12-006/PHS12-006

3. Takayuki Fujita,"Towards Data: A Human/Machine-oriented Approach of Medical Data," IHI 2012 - 2nd ACM SIGHIT

International Health Informatics Symposium,-

4. 宮川宜和，横松得滋，竹内悠，神田健介，藤田孝之，樋口行平，前中一介, "ハニカム絶縁構造を用いた低寄生容量 SOI の提

案と応用," 18th Symposium on “Microjoining and Assembly Technology in Electronics” (Mate2012),pp. 267-270

5. S. Miki, T. Fujita, T. Kotoge, Y.G. Jiang, M. Uehara, K. Kanda, K. Higuchi, K. Maenaka, "ELECTROMAGNETIC ENERGY

HARVESTER BY USING BURIED NdFeB," IEEE MEMS 2012,1221 - 1224?

6. ?秀春, 田中伸哉, 中村純, 須藤孝一, 高尾英邦, 樋口行平, 神田健介, 藤田孝之, 前中一介, "SON 技術を用いた静電容量式

絶対圧圧力センサの試作," 2012 年春期第 59 回応用物理学関係連合講演会,15p-GP3-17

7. 横松得滋, 竹内悠, 宮川宜和, 伊賀友樹, 神田健介, 藤田孝之, 樋口行平, 前中一介, "ハニカム構造体を用いた低寄生容量

SOI の検討," 2012 年春期第 59 回応用物理学関係連合講演会,16p-E3-15

8. 神田健介, 明石優, 藤田孝之, 前中一介, "低アスペクト比構造体の PZT 薄膜による側面方向駆動," 2012 年春期第 59 回応

用物理学関係連合講演会,15p-GP3-23

9. 藤田孝之, 三木省吾, 神田健介, 前中一介. 蒋永剛, 上原稔, 樋口行平, "４インチウエハ上への NdFeB 磁性薄膜の形成," 平

成 24 年電気学会全国大会講演講演論文集,3-158

10. 神田健介,永田和幸，山川隆洋，藤田孝之，前中一介, "MEMS のための薄膜 PZT/PZT バイモルフ," 第 29 回強誘電体応用会

議

11. 藤田孝之, "エナジーハーベスティング技術と材料," 2012 年日本結晶成長学会特別講演会「次世代エネルギーシステムにおけ

る結晶成長の役割」,-

12. 狩野公秀, 石井雅敏, 笠井一夫, 樋口行平, 神田健介, 藤田孝之, 前中一介,環状受光部をもつ小型脈波センサによる脈波検出,

電気学会センサ・マイクロマシン部門総合研究会フィジカルセンサ研究会,PHS-12-010

13. 橋本泰知, 牧村賢一, 宮本亜聖, 笠井一夫, 神田健介, 藤田孝之, 前中一介,垂直ミラーを用いた周回レーザ発振,電気学会セ

ンサ・マイクロマシン部門総合研究会フィジカルセンサ研究会,PHS-12-011

14. 宮本亜聖, 橋本泰知, 牧村賢一, 神田健介, 藤田孝之, 前中一介,ウエハ貼り付け技術を用いた MEMS-ガイガーカウンタの試

作,電気学会センサ・マイクロマシン部門総合研究会フィジカルセンサ研究会,PHS-12-013

15. Takayuki Fujita, Toshikazu Onishi, Kohei Fujii, Kensuke Kanda, Kazusuke Maenaka, Kohei Higuchi, "Evaluation of the

Human Vibration for Autonomous Power Source," Int Conf. World Automation Congress, 2012",

16. Toshikazu Onishi, Takayuki Fujita, Kohei Fujii, Kensuke Kanda, Kazusuke Maenaka, Kohei Higuchi, "Selective Electret

Charging Method of SiO2 Film for Energy Harvesters by Using Biased Electrode," Int Conf. World Automation

Congress, 2012",

17. Tomoya Tanaka, Takayuki Fujita, Koji Sonoda, Manabu Nii, Kensuke Kanda, Kazusuke Maenaka, Alex Chun Kit Chan,

Sayaka Okochi, Kohei Higuchi, "Wearable Health Monitoring System by Using Fuzzy Logic Heart-Rate Extraction,"

Int Conf. World Automation Congress, 2012",

18. K. Makimura, T. Hashimoto, A. Miyamoto, K. Kasai, K. Kanda, T. Fujita and K.Maenaka, "FABRICATION OF TALL

MIRRORS FOR MEMS-RLG," APCOT 2012,ac12000198

19. T. Onishi, T. Fujita, K. Fujii, K. Kanda, K. Higuchi and K. Maenaka, "ELECTROSTATIC ENERGY HARVESTER WITH

LAMINATED SIO2/SIN ELECTRET CHARGED BY BURIED GRID ELECTRODES," APCOT 2012,ac12000177

20. Y. G. Jiang, K. Kanda, T. Fujita, D. Y. Zhang and K. Maenaka, "PZT THIN FILM ACOUSTIC DEVICES FABRICATED BY

A FULL-DRY MICROMACHINING PROCESS," APCOT 2012,ac12000006

21. Alex Chun Kit Chan, Sayaka Okochi, Takayuki Fujita, Kohei Higuchi,Toshiaki Nakamura, Hirokazu Kitamura, Junichi

Kimura, and Kazusuke Maenaka, "DEVELOPMENT OF SMALL SIZED, LOW POWER AND LONG TERM OPERATING

WIRELESS SENSOR NODE," APCOT 2012,ac12000233

22. 藤田孝之, 圧電デバイスの回路・システム,GRENE 事業「低炭素社会の実現に向けた人材育成ネットワークの構築と先進環境

材料・デバイス創製」京都大学平成 24 年度夏期集中講義「創エネデバイス（圧電）コース」,-

23. K. Kanda, K. Nagata, T. Yamakawa, Y. Iga, T. Fujita, K. Maenaka, "Miniaturization of Vibratory Beam Accelerometer

by Using PZT," Eurosensors XXVI,"Vol. 47, pp.1141-1144

24. T. Fujita, S. Miki, T. Kotoge, M. Uehara, K. Kanda, K. Higuchi, K. Maenaka, "Batch fabrication technique of NdFeB for

MEMS based electromagnetic energy harvester," Eurosensors XXVI,"Vol. 47, pp. 639-642

25. 藤田孝之,JST-ERATO 前中センシング融合プロジェクトにおけるマイクロパワーデバイス,平成 24 年度エナジーハーベステ

ィングコンソーシアム総会,-

26. 藤田孝之,エナジーハーベスターの最近の研究開発動向,センサ・アクチュエータ・マイクロナノ／ウィーク 2012 次世代セン

サ総合シンポジウム,-

27. Alex Chun Kit Chan, Sayaka Okochi, Kohei Higuchi, Toshiaki Nakamura, Hirokazu Kitamura, Junichi Kimura,

Takayuki Fujita, and Kazusuke Maenaka, "Low Power Wireless Sensor Node for Human Centered Transportation

System," IEEE SMC2012, pp. 1542-1545

28. Masahide Kano, Masatoshi Ishii, Kensuke Kanda,Takayuki Fujita, Kazusuke Maenaka, Kazuo Kasai, Kohei Higuchi,

"Reflective Photoplethysmography Sensor with Ring-Shaped Photodiode," IEEE SMC2012, pp. 2058-2061

29. Yoshikazu Miyagawa, Kensuke Kanda, Takayuki Fujita, Kazusuke Maenaka, Tokuji Yokomatsu, Haruka Takeuchi,

Kohei Higuchi, "Acceleration Sensor Using SOI Wafer With Honeycomb Insulator," IEEE SMC2012, pp. 2062-2066

30. 橋本泰知, 牧村賢一, 宮本亜聖, 神田健介, 藤田孝之, 前中一介,MEMS 共振器を用いた周回レーザ発振,第 29 回「センサ・マ

イクロマシンと応用システム」シンポジウム,24-28

31. 園田晃司, 藤井孝平, 大西斗志一, 勝間洋行, 神田健介, 藤田孝之, 樋口行平, 前中一介, “振動型発電デバイスのための加速

度データ評価,” 日本機械学会第 4 回マイクロ・ナノ工学シンポジウム,OS4-1-2

32. 大西斗志一, 藤田孝之, 藤井孝平, 勝間洋行, 神田健介, 樋口行平, 前中一介, ”SiN/SiO2 高耐熱エレクトレット膜へのバイ

アス荷電手法,” 日本機械学会第 4 回マイクロ・ナノ工学シンポジウム,P-OS4-1

33. Kohei Fujii, Toshikazu Onishi, Koji Sonoda, Takayuki Fujita, Kensuke Kanda, Kohei Higuchi, and Kazusuke Maenaka,

"Development of Electret Energy Harvester with Micro Spacer," 5th International Conference on Emerging Trends in

Engineering and Technology, pp. 48-51

34. Koji Sonoda, Yuji Kishida, Tomoya Tanaka, Kensuke Kanda, Takayuki Fujita, Kazusuke Maenaka, Kohei

Higuchi",Wearable Photoplethysmographic Sensor System with PSoC Microcontroller," 5th International Conference

on Emerging Trends in Engineering and Technology, pp. 61-65

35. Asei Miyamoto, Taichi Hashimoto, Kenichi Makimura, Kensuke Kanda, Takayuki Fujita, and Kazusuke Maenaka,

"Wafer level packaging for MEMS Geiger counter," 5th International Conference on Emerging Trends in Engineering

and Technology, pp. 66-69

36. Tokuji Yokomatsu, Haruka Takeuchi, Yoshikazu Miyagawa, Kensuke Kanda, Takayuki Fujita, Kohei Higuchi, and

Kazusuke Maenaka, "Novel Honeycomb SOI Structure with Low Parasitic Capacitance for Human-Sensing

Accelerometer," 5th International Conference on Emerging Trends in Engineering and Technology, pp. 70-74

37. Takayuki Fujita, Shohei Shiono, Kensuke Kanda, Kazusuke Maenaka, Hiroyuki Hamada, Kohei Higuchi, "Flexible

Sensor for Human Monitoring System by Using P(VDF/TrFE) Thin Film," 5th International Conference on Emerging

Trends in Engineering and Technology, pp. 75-79

38. K. Kanda, Y. Akashi, T. Saito, T. Fujita, K. Higuchi, K. Maenaka, "Fabrication and Characterization of

Lateral-Directional Actuator Using Pb(Zr, Ti)O3 Thin Films," 5th International Conference on Emerging Trends in

Engineering and Technology, pp. 125-127

39. Takayuki Fujita,Electrostatic and Electromagnetic Energy Harvester for Human Monitoring System,"International

Conference on BioElectronics," BioSensors, BioMedical Devices, BioMEMS/NEMS and Applications 2012 (Bio4Apps

2012), Invite-1

40. 藤田孝之,グリーンシステム技術分科会キックオフおよびエナジーハーベスティングの現状,第 1 回グリーンシステム技術分

科会,

41. T. Fujita, T. Onishi, K. Fujii, K. Sonoda, H. Katsuma, K. Kanda, K. Higuchi, K. Maenaka, "BIPOLAR ELECTRET

CHARGING METHOD FOR ENERGY HARVESTER," PowerMEMS2012, pp. 436-439

42. 藤田孝之,PowerMEMS2012 会議報告「圧電・電磁型振動発電」,第８回マイクロエネルギー研究会,

43. 小峠竜也, 藤田孝之, 田中祐至, 上原稔, 神田健介, 樋口行平, 前中一介, "NdFeB スパッタ薄膜の微細加工と応用デバイス,"

電気学会マグネティック研究会, MAG-12-170

書籍

1. 鈴木雄二監修, 第 4 編応用第 3 章「共有電極を用いた MEMS エレクトレット発電器」藤田孝之, 環境発電ハンドブック～

電池レスワールドによる豊かな環境低負荷,NTS 出版, 2012.11

看護アプリケーショングループ

新居学助教

湯本高行助教

坂下玲子教授（看護学部）

内布敦子教授（看護学部）

Journal Papers

1. Reiko Sakashita, Atsuko Uchinuno, Kazuko Kamiizumi, Keiko Tei, and Masumi Murakami, "Nursing Quality Related to

Medical Incidents", Proc. of 2012 Fifth International Conference on Emerging Trends in Engineering & Technology,

pp.116-119, 2012


1. Manabu Nii, Yutaka Takahashi, Atsuko Uchinuno, Reiko Sakashita, "An Approach using Conceptual Fuzzy Sets for

Nursing-care Text Classification", Proc. of World Automation Congress 2012, WAC-2012-1569535307, 2012

2. Manabu Nii, Kazuki Nakai, Yutaka Takahashi, Kazusuke Maenaka, Kohei Higuchi, "Human Action Classification for a

Small Size Physical Condition Monitoring System", Proc. of World Automation Congress 2012,

WAC-2012-1569535303, 2012

3. Tomoya Tanaka, Takayuki Fujita, Koji Sonoda, Manabu Nii, Kensuke Kanda, Kazusuke Maenaka, Alex Chan Chun Kit,

Sayaka Okochi, Kohei Higuchi, "Wearable Health Monitoring System by Using Fuzzy Logic Heart-Rate Extraction",

Proc. of World Automation Congress 2012, WAC-2012-1569536039, 2012

4. Manabu Nii, Yoshihiro Kakiuchi, Tomoya Takana, Kazusuke Maenaka, Kohei Higuchi, "Implementation of a Intelligent

System into a Small Physical Condition Monitoring Device for Healthcare", Proc. of 2012 IEEE International

Conference on Systems, Man, and Cybernetics, pp.2052-2057, 2012

5. Manabu Nii, Yoshinori Hirohata, Atsuko Uchinuno, Reiko Sakashita, "New Feature Definition for Improvement of

Nursing-care Text Classification", Proc. of 2012 IEEE International Conference on Systems, Man, and Cybernetics,

pp.2610-2615, 2012

6. Manabu Nii, Yoshinori Hirohata, Atsuko Uchinuno, Reiko Sakashita, "Feature Definition using Dependency Relations

between Terms for Improving Nursing-care Text Classification", Proc. of 2012 Fifth International Conference on

Emerging Trends in Engineering & Technology, pp.110-115, 2012

7. Takuo Hirayama, Takayuki Yumoto, Manabu Nii and Kunihiro Sato, “A System for Visualization and Navigation of

Customer Reviews”, IADIS International Conference Internet Technologies & Society 2012, pp.259-264 (2012).

8. 湯本高行，“ハウツー情報間の共通性に基づく検索結果の提示手法”, 電子情報通信学会第 22 回 Web インテリジェンスとイ

ンタラクション研究会，pp.83-84 (2012).

9. 平山拓央, 湯本高行, 新居学, 佐藤邦弘, “語の共起と極性に基づく商品レビュー閲覧支援システム” , 第 155 回情報処理

学会データベースシステム研究発表会, 2012-DBS-155(3), pp.1-9 (2012).

10. 多田亮平，湯本高行，新居学，佐藤邦弘, “カテゴリに対する所属度と典型度を考慮した希少な Web ページの発見”, 第

155 回情報処理学会データベースシステム研究発表会, 2012-DBS-155(13), pp.1-8 (2012).

生体計測シミュレーショングループ

倉本圭准教授

林治尚准教授

下權谷祐児助教

齋藤歩助教

比嘉昌助教

佐藤哲也教授（シミュレーション学研究科）

内海裕一教授（高度産業科学研究所）

Journal Papers

1. Y. Shimogonya, H. Kumamaru, and K. Itoh, “Sensitivity of the gradient oscillatory number to flow input waveform

shapes,” Journal of Biomechanics, Vol.45, pp.985-989, 2012.

2. A. Saitoh, K. Miyashita, T. Itoh, A. Kamitani, T. Isokawa, N. Kamiura and N. Matsui, “Accuracy Improvement of

Extended Boundary-Node Method” to be published in IEEE Trans. Magn.

3. A. Saitoh, T. Itoh, N. Matsui, A. Kamitani and H. Nakamura, “Application of Collocation Meshless Method to

Eigenvalue Problem,” Plasma Fusion Res., Vol. 7, 2406096, 2012.

4. T. Itoh, A. Saitoh, A. Kamitani and H. Nakamura, “Implicit Function with a Natural Behavior over the Entire Domain,”

Plasma Fusion Res., Vol. 7, 2406068, 2012.

5. A. Kamitani, T. Takayama, A. Saitoh and H. Nakamura, “Accurate and Stable Numerical Method for Analyzing

Shielding Current Density in High-Temperature Superconducting Film Containing Cracks,” Plasma Fusion Res., Vol.

7, 2405024, 2012.

6. T. Takayama, A. Kamitani, A. Saitoh and H. Nakamura, “Numerical Investigation on Accuracy and Resolution of

Contactless Methods for Measuring jC in High-Temperature Superconducting Film: Inductive Method and

Permanent Magnet Method,” Plasma Fusion Res., Vol. 7, 2405017, 2012.

7. T. Takayama, A. Saitoh and A. Kamitani, “Numerical Simulation of Conventional/Enhanced Permanent Magnet

Method: Influence of Crack on Accuracy,” IEEE Trans. Appl. Supercond., Vol. 22, 4903904, 2012.


1. Y. Shimogonya, H. Kumamaru, and K. Itoh, “Hemodynamic in cerebral arteries before aneurysm formation:

Influence of flow input waveform shapes,” 8th European Solid Mechanics Conference, MS-24.2, 2012.

2. H. Kumamaru, K. Itoh, and Y. Shimogonya, “Numerical analyses on liquid-metal magnetohydrodynamic flow in

180-degrees u-bend with small curvature radius,” International Computational Mechanics Symposium 2012,

MS7-2-2, 2012.

3. Y. Kaku, K. Kuramoto, S. Kobashi and Y. Hata, "Predict time series data for the number of asthmatic attacks in Himeji

by Fuzzy-AR model," Proc. of 2012 Fifth Int. Conf. on Emerging Trends in Engineering and Technology, pp. 314-317,

2012.

4. T. Ohkawa, S. Kobashi, Y. Hata, Y. Yamashita, N. Aono and K. Kuramoto, "Theoretical examination of carrier

interaction of Pd catalyst," Proc. of 7th Int. Conf. on Environmental Catalysis, 2012.

5. Y. Kaku, K. Kuramoto, S. Kobashi and Y. Hata, "Asthmatic attacks prediction considering weather factors based on

Fuzzy-AR model," Proc. of 2012 IEEE World Congress on Computational Intelligence, pp. 2023-2026, 2012.

6. 福田俊一, 下權谷祐児, 塚原徹也, 青木友和, 川端康弘, 松井雄哉, 井本恭秀, 福田美雪, “脳動脈瘤 CFD 解析における流入条

件の設定について〜既成値と実測値での相違〜,” 第 38 回日本脳卒中学会総会, SS-O05-4, 2013.

7. 井本恭秀, 下權谷祐児, 福田俊一, 熊丸博滋, 伊藤和宏, “脳動脈瘤実症例データに基づく個体別モデルの構築と瘤発生要因

の検討への応用,” 日本機械学会関西学生会平成 24 年度卒業研究発表講演会前刷集, 7A25, 2013.

8. 下權谷祐児, 西野修平, 熊丸博滋, 伊藤和宏, “拍動流中に投与された微粒子の磁気送達に関する基礎的研究,” 計算工学講演

会論文集, Vol.17, F-5-4, 2012.

9. 下權谷祐児, 福田俊一, 井本恭秀, 里啓太郎, 熊丸博滋, 伊藤和宏, “脳動脈瘤形成部位における血行力学的負荷の評価：流入

条件の影響,” 日本機械学会第 25 回バイオエンジニアリング講演会講演論文集, pp.397-398, 2012.

10. 里啓太郎, 下權谷祐児, 熊丸博滋, 伊藤和宏, “流入条件の違いが脳動脈瘤形成部位における血行力学的負荷に及ぼす影響,”

可視化情報全国講演会（姫路 2012）講演論文集, pp.145-146, 2012.

11. 熊丸博滋, 大上晶也, 阪田風馬, 伊藤和宏, 下權谷祐児, “磁性体球回転式マイクロポンプに関する研究,” 日本機械学会

2012 年度年次大会講演論文集, JS53052, 2012.

12. 熊丸博滋, 伊藤和宏, 下權谷祐児, “急拡大流路内の液体金属電磁流体流れに関する数値解析（印加磁場に垂直な方向に急拡大

する場合）,” 日本機械学会第 90 期流体工学部門講演会講演論文集, G602, 2012.

13. 尾辻達志，三枝弘幸，船橋昇広，嶺重温，畑豊，倉本圭， "アパタイト型イオン伝導結晶のラマン分光に関する研究,"

第 2 回量子化学ウィンタースクール, 2012.

14. 中村篤史，小橋昌司，畑豊，倉本圭， "ファジィ理論を用いた光電子スペクトルピーク予測の高精度化," 第 2 回量子化

学ウィンタースクール, 2012.

15. 大川哲也，山下嘉典，青野紀彦，小橋昌司，畑豊，倉本圭， "Pd 担持触媒を用いた自動車排ガス浄化に関する理論的研

究," 第 2 回量子化学ウィンタースクール, 2012.

16. 倉本圭， "触媒表面分析計測技術と情報科学の接点," 第 110 回触媒討論会, 2012.（依頼公演）

17. 大川哲也，小橋昌司，畑豊，山下嘉典，青野紀彦，倉本圭， "第一原理熱力学計算による Pd 触媒を用いた CO 酸化反応

における浄化温度の理論的予測," 第 110 回触媒討論会, 2012.

18. 田中基史，鈴木春洋，小橋昌司，畑豊，倉本圭， "パターン投影法を用いた非接触三次元計測の触媒表面分析への応用,"

第 110 回触媒討論会, 2012.

19. 平井公貴，小橋昌司，畑豊，倉本圭， "多成分の金属種を含むクラスターモデルを用いた NO 還元反応の量子化学計算,"

第 110 回触媒討論会, 2012.

20. 西浦孝則，才木常正，小橋昌司，畑豊，内海裕一，倉本圭， "弾性表面波を用いた粉末輸送機構の粒子法によるシミュレーシ

ョン," 第 110 回触媒討論会, 2012.

21. 郭悠翔，菊池翔，田中基史，倉本圭， "姫路市の天気・医療情報を活用した喘息患者数予測システム開発と応用〜姫路

市と連携した政策研究と健康増進〜," 兵庫県立大学研究テーマ集セレクション 2012，p.22, 2012.

22. 郭悠翔，小橋昌司，倉本圭，畑豊， "Fuzzy-AR モデルを用いた姫路市の年代別喘息発作数予測," 第 28 回ファジィシ

ステムシンポジウム講演論文集, pp. 193-198, 2012.

23. 菊池翔，郭悠翔，小橋昌司，倉本圭，畑豊， "AR モデルを用いた姫路市の地域別喘息発作数予測," 第 28 回ファジ

ィシステムシンポジウム講演論文集, pp. 203-206, 2012.

24. 中村篤史，倉本圭，小橋昌司，畑豊， "含窒素分子に対するファジィ推論を用いた光電子スペクトルピーク予測," 第 28

回ファジィシステムシンポジウム講演論文集, pp. 187-192, 2012.

25. 中村篤史，倉本圭，小橋昌司，畑豊 "ファジィ推論を用いた光電子スペクトルピーク予測の高精度化," 多値論理研究ノ

ート, Vol. 35, No. 2, pp. 1-6, 2012.

26. 田中基史，倉本圭，小橋昌司，畑豊 "ファジィ推論を用いたパターン投影法による非接触三次元計測の高精度化," 多値

論理研究ノート, Vol. 35, No. 3, pp. 1-6, 2012.

27. 郭悠翔，倉本圭，小橋昌司，畑豊 "Fuzzy-AR モデルを用いた喘息発作数予測," 多値論理研究ノート, Vol. 35, No. 21,

pp. 1-6, 2012.

28. 郭悠翔, 倉本圭, 小橋昌司, 畑豊, "姫路市の喘息発作に関する研究," 情報通信学会総合大会２０１２講演論文集,

D-7-1, 2012.

29. 宇野健, 佐々和洋, 林治尚, 中野英彦, “Flash を用いた分子グラフィックスシステムの AR への応用,” 日本コンピュータ化

学会 2012 秋季年会, 2012.

30. 佐々和洋, 長井昭太朗, 林治尚, 後藤仁志, “DPCミセルの分子動力学的基礎解析,” 日本コンピュータ化学会 2012秋季年会,

2012.

31. 林治尚, 畑豊, 太田勲, “兵庫県立大学における法人化に向けた情報関連システムの改編,” 大学 ICT推進協議会 2012年度

年次大会, 2012.

32. Higa, M; Struk, A.M.; Roche, C.; Farmer, K.W.; Wright, T.; Banks, S.A., “An instrumented trial prosthesis for

intraoperative joint force measurements during reverse total shoulder arthroplasty”, Transaction of the 59th

Orthopaedic research Society, p. 1891

33. Ito, H; Tanino, H; Yamanaka, Y; Sato, T; Nishida, Y;1Nakamura, T; Higa, M, “Decrease in Articular Cartilage

Thickness of a Loaded Hip Joint Before and After Excision of the Labrum”, Transaction of the 59th Orthopaedic

research Society, p. 971

34. Masaru Higa, kazumasa Matsuda, Masayoshi Abo, Satoshi Kakunai, “Measurements of passive resisting moment of

the hip in younger and elder subjects,” 第 51 回日本生体医工学学会

35. K. Miyashita, A. Saitoh, T. Itoh, A. Kamitani, N. Kamiura, and N. Matsui, “Development of Modified Extended

Boundary-Node Method -New Approach for Determining Data Points-,” Proceedings of JSST 2012, pp. 112-115,

2012.

36. M. Fujii, T. Isokawa, H. Ikeno, N. Kamiura, A. Saitoh, and N. Matsui, “On Detecting Waggle Dance Behaviors from

Spatio-Temporal Sectional Images,” Proceedings of SICE Annual Conf. 2012, pp. 439-442, 2012.

Chapter in Book

1. Y. Shimogonya, H. Kumamaru, and K. Itoh, “Computational study of the hemodynamics of cerebral aneurysm

initiation,” Biomedical Engineering and Cognitive Neuroscience for Healthcare: Interdisciplinary Applications,

pp.267-277, 2012.

2. H. Kumamaru, K. Itoh, and Y. Shimogonya, “Three-dimensional numerical analyses on liquid-metal

magnetohydrodynamic flow through circular pipe in magnetic-field outlet-region,” Trends in Electromagnetism –

From Fundamentals to Applications, pp.207-222, 2012.

ヘルスケアシステムグループ

相河聡教授

畠山賢一教授

河合正准教授

山本真一郎助教

内田勇人教授（環境人間学部）

Journal Papers

1. 峰松岳志，相河聡， “ロボットの位置推定・移動履歴を用いた到達精度向上に関する実験的検討，”電子情報通信学会論文誌

B, Vol.J95-B, No.9, pp.1218-1222, 2012.

2. 林大貴，相河聡，“モバイル無線 LAN ルータのスループット向上のための動的チャネル選択手法の検討，”電子情報通信学会論

文誌 B, Vol.J95-B, No.8, pp.980-984, 2012.

3. S. Yamamoto, D. Ishihara and K. Hatakeyama，“Design of Electromagnetic Wave Absorber Panels for Oblique

Incidence Using Wire Array Sheet“，IEICE Transactions on Communications，Vol.E95-B，No.2，pp.631-634，2012.

4. 岩井通，山本真一郎，畠山賢一，“金属格子と誘電体積層構造の反射透過特性，”電子情報通信学会論文誌 B，Vol. J95-B，No.

3，pp. 488-492，2012.

5. 岩井通，山本真一郎，畠山賢一，“金属格子－誘電体－金属格子積層構造の電磁波透過特性，” 電気学会論文誌 A，Vol. 132，

No. 5，pp. 385-386，2012.

6. 山本真一郎，石原大輔，畠山賢一，“二つの共鳴周波数を有する金属線配列材を用いた広帯域整合型電波吸収体の設計・評価，”

電子情報通信学会論文誌 B，Vol. J95-B，No. 6，pp. 734-741，2012.

7. 桜井良太，藤原佳典，深谷太郎，齋藤京子，安永正史，鈴木宏幸，野中久美子，金憲経，金美芝，田中千晶，西川武志，内田

勇人，新開省二，渡辺修一郎. 運動に対する充足感が高齢者および高齢者の運動介入効果に与える影響－運動充足感と身体活

動量からの検討－，日本公衆衛生雑誌，59(10)，743-753，2012 年 10 月

8. 濵口郁枝，奥田豊子，内田勇人，大喜多祥子，福本タミ子，北元憲利．大学生に対する食育が食行動に及ぼす影響，日本食育

学会誌，6(3)，257-264，2012 年 7 月

9. 濵口郁枝，奥田豊子，内田勇人，大喜多祥子，福本タミ子，北元憲利．大学生に対する食育の効果の検証，日本食育学会誌，

6(3)，249-255，2012 年 7 月

10. 田中千晶，藤原佳典，安永正史，桜井良太，斎藤京子，金憲経，深谷太郎，野中久美子，小林和成，吉田裕人，内田勇人，

新開省二，渡辺修一郎．複合健康増進プログラムが地域在住高齢者の日常的な身体活動量へ与える影響-無作為化比較試験に

よる検討-，日本老年医学会雑誌，49(3)，372-374，2012 年 5 月

11. 内田勇人，藤原佳典，西垣利男，香川雅春，作田はるみ，下村尚美，木宮高代，濱口郁枝，東根裕子，矢野真理．高齢者によ

る育児支援活動が高齢者の心身の健康と母親の育児ストレスへ及ぼす影響，日本世代間交流学会誌，2(1)，

12. 33-39，2012 年 3 月

13. 濱口郁枝, 内田勇人, 奥田豊子, 作田はるみ, 大喜多祥子, 福本タミ子, 北元憲利．女子大学生に対する味覚教育の実施が味

覚能力に及ぼす影響，小児保健研究，71(2)，304-315，2012 年 3 月


1. 坂口大樹,相河聡,“減衰定数を学習する位置推定アルゴリズム，” 電子情報通信学会, vol. 112, no. 334, CS2012-82,

IE2012-96, pp. 65-70, 2012.

2. 坂口大樹,相河聡,“無線 LAN 端末の位置・精度情報を用いた重み付け位置推定法の実験的検討，” 2012 年電子情報通信学会

通信ソサイエティ大会, B-19-13, 2012.

3. 大川原健太，西垣剛士，相河聡,“MIMO アンテナを用いた秘密鍵共有方式における秘密鍵生成高速化に関する検討，電子

情報通信学会関西支部学生会, 第 18 回学生会研究発表講演会, B3-3, 2013.3.

4. 馬場亮輔，相河聡,“無線 LAN 端末を用いた位置推定の Scene Analysis 方式におけるデータベース平滑化・内挿について

の検討，” 電子情報通信学会関西支部学生会, 第 18 回学生会研究発表講演会, B4-1, 2013.3.

5. 重松祐平，相河聡, “屋内屋外の位置推定手法切り替えに関する研究，” 電子情報通信学会関西支部学生会, 第 18 回学生会

研究発表講演会, B4-2, 2013.3.

6. 西垣剛士，相河聡, “伝搬路特性を用いた秘密鍵の 2 地点受信による盗聴法に関する検討，” 2013 年電子情報通信学会総合大

会, B-5-144, 2013.3.

7. 林祐一郎，相河聡, “複数アンテナを用いた高精度位置推定法の実験的検討，”2013 年電子情報通信学会総合大会, B-19-43,

2013.3.

8. S. Yamamoto，D. Ishihara，K. Hatakeyama，“(Invited-paper) EM Absorber and Shielding Screen Using Metal Wire

Array Sheet，”Proceedings，2012 Korea-Japan EMT/EMC/BE Joint Conference，KJJC-2012，pp.221-224，2012.

9. S. Yamamoto，K. Hatakeyama，“(Invited-paper) Design Method of EM Absorber and Shielding Screen Using Wire

Array Sheet，”Proceedings，2012 IEEE International Symposium on Electromagnetic Compatibility，pp.416-421，

2012.

10. T. Iwai，S. Yamamoto，S. Aikawa，K. Hatakeyama，“Electro-magnetic Shielding Technologies in Hospital，”Proceedings，

2012 Fifth International Conference on Emerging Trends in Engineering and Technology，ICETET-12，pp.321-324，

2012.

11. 山本真一郎，岩井通，畠山賢一，“人工材料を用いる反射・透過制御材の設計法，”電子情報通信学会総合大会，通信環境の制

御に関わるメタマテリアルの要素技術 BI-2-4，pp.SS-69-SS-70，2012.

12. 長谷川和明，岩井通，山本真一郎，畠山賢一，“ミリ波帯用電波吸収体の斜め入射特性測定例，”信学技報，EMCJ2011-119，

pp.47-52，2012.

13. 橋本流一，岩井通，山本真一郎，畠山賢一，“金属線配列材と 1/4 波長整合層による広帯域電波吸収体，”信学技報，

EMCJ2011-120，pp.53-58，2012.

14. 末﨑健太，岩井通，山本真一郎，畠山賢一，“金属格子と誘電体材料を用いた反射・透過制御材，” 信学会総合大会，B-4-50，

p.380，2012.

15. 佐藤毅博，山下意仁，山本真一郎，中村龍哉，畠山賢一，廣瀬美佳，葭内暁，“実用化に向けた燻し瓦ピラミッド形電波吸収

体の設計・評価，” 信学会総合大会，B-4-51，p.381，2012.

16. 佐藤毅博，山本真一郎，中村龍哉，畠山賢一，廣瀬美佳，葭内暁，“燻し瓦ピラミッド形電波吸収体の水膜および汚れの影響

に関する検討，”信学会ソサイエティ大会，B-4-59，p.355，2012.

17. 末﨑健太，山本真一郎，畠山賢一，岩井通，“金属格子と誘電体積層した空間高域通過フィルター，”信学会ソサイエティ大会，

B-4-67，p.363，2012.

18. 薄木誠，畠山賢一，山本真一郎，“金属線配列材の低周波帯誘電率について，”平成 24 年電気関係学会関西支部連合大会，

8amD-11，pp.66-67，2012.

19. 小寺達明，山本真一郎，畠山賢一，“ミリ波帯斜め入射特性測定装置を用いた電波吸収特性評価，”平成 24 年電気関係学会関西

支部連合大会，P-35，pp.535-536，2012.

20. 小寺達明，山本真一郎，畠山賢一，岩井通，“40GHz から 75GHz に渡るミリ波用電波吸収体の斜め入射特性評価，”信学技

報，EMCJ2012-116，pp.81-86，2013.

21. 北澤広行，薄木誠，山本真一郎，畠山賢一，“単共鳴形・二重共鳴形金属線配列材の電磁遮蔽特性評価，”電子情報通信学会関

西支部，第 18 回学生会研究発表講演会，B1-1，p.21，2013.

22. 有川大三，佐藤毅博，山本真一郎，中村龍哉，畠山賢一，“燻し瓦製造法によるピラミッド形電波吸収体の水膜付着による影響，”

電子情報通信学会関西支部，第 18 回学生会研究発表講演会，B7-6，p.60，2013.

23. 薄木誠，山本真一郎，畠山賢一，“二つの共鳴周波数を有する金属線配列材の簡易比誘電率推定法，” 信学会総合大会，B-4-4，

p.349，2013.

24. Hayato Uchida, Fujio Hirata, Nobuyuki Mino, Masahiro Toyoda, Masahiro Yamashita, Toshimitsu Tachibana, Toshio

Nishigaki, Hideki Toji. The related factors of frailty risk in Japanese Elderly Women, Proceedings of World

Automation Congress 2012, Puerto Vallarta, Mexico. 2012.6

25. Hayato Uchida, Yoshinori Fujiwara, Shoji Shinkai, Toshio Nishigaki, Yukari Yamashita, Yuko Hori, Hideki Toji,

Yoshihiro Hasegawa, The Changes of Frailty Risk in Japanese Elderly Women and its Related Factors, 65th Annual

Scientific Meeting of Gerontological Society of America, San Diego, USA. 2012.11

26. Hayato Uchida, Toshio Nishigaki, Hideki Toji, Mari Yano, Jiro Kuroda, Teiji Okamoto, Yoshinori Fujiwara. The risk of

developing a need for long-term care and the associated factors in Japanese elderly women, 8th World Congress on

Active Ageing, Glasgow, UK. 2012.8

27. 矢野真理, 内田勇人, 森山知恵, 藤原佳典，地域在住高齢女性における介護予防に関する縦断的研究外出行動に着目し

て，第 71 回日本公衆衛生学会総会，2012 年 10 月，山口

28. 森山知恵, 内田勇人, 矢野真理, 藤原佳典，地域在住高齢女性における介護予防に関する縦断的研究抑うつに着目して，

第 71 回日本公衆衛生学会総会，2012 年 10 月，山口

29. 鈴木宏幸, 佐久間尚子, 安永正史, 大場宏美, 小宇佐陽子, 野中久美子, 石井賢二, 内田勇人, 新開省二, 藤原佳

典，世代間交流プログラム REPRINTS の 5 年間の縦断的検討言語流暢性への効果，第 71 回日本公衆衛生学会総会，2012

年 10 月，山口

30. 佐久間尚子, 鈴木宏幸, 安永正史, 大場宏美, 小宇佐陽子, 野中久美子, 石井賢二, 内田勇人, 新開省二, 藤原佳

典，世代間交流プログラム REPRINTS の 5 年間の縦断的検討高齢者の認知機能の側面，第 71 回日本公衆衛生学会総会，

2012 年 10 月，山口

31. 内田勇人, 藤原佳典, 森山知恵, 矢野真理, 松浦伸郎，高齢者による自然体験活動支援が児童養護施設入所児童の高齢者

イメージに及ぼす影響，第 71 回日本公衆衛生学会総会，2012 年 10 月，山口

32. 松浦伸郎, 内田勇人，学校における児童・生徒の体格追跡調査，第 115 回日本小児科学会学術集会，2012 年 3 月

33. Jingfan Xu, Tadashi Kawai, Member, IEEE, Akira Enokihara, Member, IEEE,

34. and Hiroshi MurakamiInvestigation of Heating Effect by Microwave Irradiation for Phantoms and Plants

著書・解説

1. 内田勇人．『多様化社会をつむぐ世代間交流－次世代への『いのち』の連鎖をつなぐ－（草野篤子、内田勇人、溝邊和成、吉

津晶子編）』, 三学出版（滋賀）．第 12 章「シニアボランティアによる小学校教育支援活動」を担当．2012 年 7 月

2. 内田勇人．『スポーツビジネス概論』（黒田次郎・遠藤利文編集），叢文社（東京），25 名で分担執筆．第 10 章「医療・介護

関連の市場（pp.111-122）」を担当．2012 年 3 月

3. 内田勇人．シックハウス症候群．おかしいなと思ったらすぐ換気．１日 30 分の空気の入れ替えで症状改善を，播磨リビング

新聞（女性と子どものための医療ガイド、- Vol.16），2012 年 1 月

その他

1. 平成２４年度電波の日近畿総合通信局長表彰(個人)H24.6.1

（SHF 帯を活用した地上デジタル放送配信システムの実用化に向けて技術的課題を検証するなど、放送の技術開発に多大な

貢献をされました。）

2. 電波新聞５面ＳＨＦ帯地デジ放送難視聴対策、実用化に期待, 2012.4.4.

生体数理モデリンググループ

松井伸之教授（副センター長）

上浦尚武准教授

礒川悌次郎准教授

池野英利教授（環境人間学部）


1. N.Kamiura, A.Saitoh, T.Isokawa, N.Matsui, and H.Tabuchi, "On classification of interview sheets for ophthalmic

examinations using self-organizing maps," Proceedings of the 17th International Conference on Artificial Life and

Robotics (AROB 17th '12), pp.686-689, 2012.

2. M.Fujii, T.Isokawa, H.Ikeno, N.Kamiura, A.Saitoh, and N.Matsui, "On Detecting Waggle Dance Behaviors from

Spatio-Temporal Sectional Images," Proceedings of SICE Annual Conference 2012, pp.439-442, 2012.

3. N.Muramoto, N.Matsui, and T.Isokawa, "Searching Ability of Qubit-Inspired Genetic Algorithm," Proceedings of SICE

Annual Conference 2012, pp.443-446, 2012.

4. N.Kamiura, A.Saitoh, T.Isokawa, N.Matsui, and H.Tabuchi, "On Waiting-time Estimation for Outpatients at

Department of Ophthalmology," Proceedings of Fifth International Conference on Emerging Trends in Engineering

and Technology (ICETET-12), pp.24-29, 2012.

5. K.Omachi, T.Isokawa, N.Kamiura, H.Nishimura, and N.Matsui, "Retrieval Performance of Complex-Valued

Associative Memory with Complex Network Structure," Proceedings of Fifth International Conference on Emerging

Trends in Engineering and Technology (ICETET-12), pp.40-43, 2012.

6. N.Kamiura, A.Saitoh, T.Isokawa, N.Matsui, and H.Tabuchi, "Classification of Interview Sheets Using Self-Organizing

Maps for Determination of Ophthalmic Examinations," Lecture Notes in Computer Science, vol.7666, pp.148-155,

2012.

7. T.Isokawa, H.Nishimura, and N.Matsui, "Quaternionic Multilayer Perceptron with Local Analyticity," Information,

vol.3, no.4, pp.756-770, 2012.

8. 上浦尚武, 齋藤歩, 礒川悌次郎, 松井伸之, 田淵仁志, "眼科診療における待ち時間予測に関する一考察," 多値論理研究ノー

ト, vol.35, no.19, pp.19-1-19-6 (2012)

9. 田附浩一朗, 村本憲幸, 松井伸之, 礒川悌次郎, "量子粒子群最適化法の基本性能評価," 第 2 回コンピューテーショナル・イ

ンテリジェンス研究会講演論文集, pp.121-124 (2012)

10. 村本憲幸, 田附浩一朗, 松井伸之, 礒川悌次郎, "非定常環境下における量子粒子群最適化法の性能評価," 計測自動制御学会

システム・情報部門学術講演会論文集, pp.140-143 (2012)

11. 大町耕平, 礒川悌次郎, 西村治彦, 上浦尚武, 松井伸之, "複雑ネットワーク構造を有する複素連想記憶システムにおける想起

性能評価," 計測自動制御学会システム・情報部門学術講演会論文集, pp.149-152 (2012)

研究者名資金名課名

荒木望科学研究費若手研究(B) 表面筋電信号を用いた関節角度推定型筋電義手の開発

礒川悌次郎科学研究費若手研究(B) 多元数表現に基づくニューラルネットワークによる幾何学的情

報処理機構の研究

内田勇人科学研究費基盤研究(C) 世代間交流事業が児童養護施設入所児童と高齢者の心身の健康

に及ぼす影響

受託研究姫路市「介護予防支援活動研究事業」

小橋昌司科学研究費若手研究(A) 知的医用画像処理による子どもの頭部 MR 画像からの難治てん

かん原性域の自動検出

齊藤歩科学研究費基盤研究(B) 有限節点法,境界節点法の完全メッシュレス化とその工学的応用

佐藤邦弘科学研究費基盤研究(B) 電子ホログラフィによる実時間３次元撮像・３次元ディスプレ

イシステムの開発

科学研究費挑戦的萌芽生体観察のためのホログラフィック４次元顕微鏡の開発

下權谷祐児科学研究費若手研究(B) 脳動脈瘤血管内治療の術前力学シミュレーション手法の開発と

実症例モデルへの応用

森本雅和経済産業省戦略的基盤

技術高度化支援事業

パン等の画像識別による POS システム組込みソフトウェアの

開発

湯本高行科学研究費若手研究(B) 情報の詳細関係に基づく Web ページの組織化

http://jp.diginfo.tv/v/12-0145-r-jp.php

兵庫県立大学大学院工学研究科医療健康情報技術研究センター

Himeji Initiative in Computational Medical and Health Technology

平成２４年度年次報告（平成２５年３月）

発行兵庫県立大学大学院工学研究科医療健康情報技術研究センター

〒671-2280 兵庫県姫路市書写 2167

TEL: 079-266-1661（代表） FAX: 079-266-8868

http://www.eng.u-hyogo.ac.jp/eecs/hata/hi-med/index.html

印刷小野高速印刷株式会社

〒670-0933 姫路市平野町 62 番地

TEL 079-281-0008 FAX 079-223-3523

Documents

Annual Report 2012 Himeji Initiative...Research Center, Osaka University, Osaka, and Himeji Initiative in Computational Medical and Health Technology, Hyogo, Japan. al. have successfully