Upload
jack-chan
View
219
Download
0
Embed Size (px)
Citation preview
8/9/2019 Mandani
1/8
A COMPARISON OF MANDANI AND SUGENO INFERENCE SYSTEMS FOR A
SPACE FAULT DETECTION APPLICATION
J. J. Jassbi, Azad University, Science and Research Campus, Iran, [email protected]
Paulo J. A. Serra, UNINOVA, Portugal, [email protected]
Rita A. Ribeiro, UNINOVA, Portugal, rar@ uninova.pt
Alessandro Donati, ESA/ESOC, Darmstad, Germany, [email protected]
ABSTRACT
This research provides a comparison between the performances of TSK (Takagi, Sugeno, Kang)-type versus Mamdani-type fuzzy inference systems. The main motivation behind this research
was to assess which approach provides the best performance for a gyroscope fault-detection
application, developed in 2002 for the European Space Agency (ESA) satellite ENVISAT. Due to
the importance of performance in online systems we compare the application, developed with
Mamdani model, with a TSK formulation using three types of tests: processing time for both
systems, robustness in the presence of randomly generated noise; and sensitivity analysis of the
systems behaviors to changes in input data. The results show that the TSK model perform better
in all three tests, hence we may conclude that replacing a Mamdani system with an equivalent
TSK system could be a good option to improve the overall performance of a fuzzy inference
system.
KEYWORDS: Fuzzy Inference Systems, Fault detection, Sensitivity analysis
1. INTRODUCTION
Intelligent monitoring and diagnostic tools are essential for mission control purposes, as for
example assessing and monitoring the health status of spacecraft and satellite components, such as
gyroscopes. In 2002 UNINOVA (Portugal) in cooperation with GTD (Spain), developed a
software system called ENVISAT GYROSCOPE MONITOR for the European Space Agency
(ESA) (see details in [1]). The fuzzy inference system (FIS) developed in the ESA project for
monitoring the gyroscopes of the ENVISAT satellite provides the spacecraft controller's team,
responsible for satellite operations, with a new type of gyroscope health-monitoring tool. This
monitoring tool generates alarms with different degrees of criticality and a severity level of the
alarm itself, instead of a simple presence/absence of an alarm. It also supplies an explanation forthe triggering of the alarm. The tool includes three main components, all developed with a
Mamdani inference mechanism [1], [2]: (1) a fault detection module that monitors the health of
each of the 4 gyroscopes; (2) a data quality system to assess the quality of the data triggering the
alarms; (3) and a generic fault detection module for the overall system.
In this paper we only use the last case, generic system fault detection module, to explain the
comparison between both Mamdani [3] and TSK (Takagi, Sugeno and Kang) FIS [4] [5] and to
discuss the performance tests made. The main objective behind this research was to assess which
approach provides the best performance for the Space Fault Detection application. Due to the
importance of inference time and robustness of results in online systems, we compare the
application, developed with Mamdani model, with a TSK formulation using three tests:
processing time, i.e. operational process time of each model in the same situation; robustness of
8/9/2019 Mandani
2/8
the system in case of noise; and sensitivity analysis of the system behavior when we change the
input variables from its minimum values to its maximum, and from zero to max (variables can be
seen in Table 2).
The paper is organized as follows. The next section presents the motivation for the comparison ofMamdani versus TSK types of inference systems. In section 3 the case study and variables are
introduced. In section 4 the results of comparing the performance of the two inference schemes
are discussed and finally in section 5 we present the conclusions.
2. MOTIVATION FOR COMPARING MANDANI FIS AND TSK FIS
In terms of inference process there are two main types of Fuzzy Inference Systems (FIS): the
Mamdani-type [3] and the TSK-type [4].
In terms of use, the Mamdani FIS is more widely used, mostly because it provides reasonable
results with a relatively simple structure, and also due to the intuitive and interpretable nature of
the rule base. Since the consequents of the rules in a TSK FIS are not fuzzy this interpretability is
lost; however, since the TSK FISs rules consequents can have as many parameters per rule asinput values, this translates into more degrees of freedom in the design than a Mamdani FIS thus
providing the systems designer with more flexibility in the design of the system [5]. However, it
should be noted that the Mamdani FIS can be used directly for either MISO systems (multiple
input single output) as well as for MIMO systems (multiple input multiple output), while the TSK
FIS can only be used in MISO systems (we explain below this issue).
The fact that a Mamdani FIS can be seen as a function that maps the systems input space into its
output space, ensures that there exists a TSK FIS that can approximate any given Mamdani FIS
with an arbitrary level of precision (based on a result known as universal approximation theorem
[6]). It is beyond the scope in this paper to explain in detail the formalisms of this comparison.
For a comprehensive comparison and description on several approximate reasoning methods,
including Mamdani FISs and TSK FISs, see [7].Summarizing, our main motivations for testing the Space monitoring application developed with
Mamdani FIS and with a TSK FIS and to compare the results are:
The TSK FIS is more flexible because it allows more parameters in the output and since
the output is a function of the inputs it expresses a more explicit relation among them;
In computational terms the TSK FIS is more effective because the complex
defuzzification process of the Mamdani FIS is replaced with a weighted average;
Because of the structure of the TSK FIS rule outputs, it is more adequate for functional
analysis than a Mamdani FIS.
From the above, it seems that any TSK FIS is always more efficient than a Mamdani FIS and the
question to ask is why wasnt the Space monitoring application developed from scratch with a
TSK FIS? There are two important reasons for this:
1. For classification problems, where the rules outputs are usually independent of each other,
i.e. MIMO systems, it does not make any sense to aggregate different nature outputs with
a weighted average. However, to select the output with the best match (max-min
inference), as in Mamdani FIS, makes perfect sense. TSK FIS is ONLYsuitable for MISO
problems, i.e. systems with the same output linguistic variable. Of course, any MIMO can
always be divided into several MISOs
2. The monitoring tool developed included two MISO FIS (the gyroscopes fault detection
and the data quality fault detection) but only the generic system level of the alarms is a
MISO system. Hence, we selected the latter to develop all modules with the same type of
FIS.
8/9/2019 Mandani
3/8
In summary, in this research only the generic system level alarms module is considered for the
performance comparison of the two FIS because it is a MISO and the three tests performed with
both FIS are: processing time, robustness with respect to noise and sensitivity analysis to changes
in the inputs.
3. GYROSCOPES GENERIC SYSTEM LEVEL
Fig.1. shows the general system level model (from the ESA monitoring tool) which is used in this
work for comparing the Mamdani and TSK FIS. Please note that some of the input variables, in
the left side, were divided in two (as shown in the Mamdani FIS box- Fig 1), thus resulting in 11
input variables in the generic system level monitoring system.
Theta_X
Theta_Y
Theta_Z
Drift_X
Drift_Y
Drift_Z
Incoherency Index
Max
Max
Max
Max/Max Delta
Max/Max Delta
Max/Max Delta
Max
Count
MANDANI FUZZY SYSTEM
Max_Theta_X
Max_Theta_Y
Max_Theta_Z
Max_Drift_X
Max_Drift_Y
Max_Drift_Z
Max_Incoherency Index
Max_Delta_Drift_X
Max_Delta_Drift_Y
Max_Delta_Drift_Z
Times_Exceds_Incoherency
No Alarm
No-Gyro Soft Alert
Soft Gyro Alert
Gyro Alert
Serious Gyro Alert
Critical Gyro Alert
Figure 1. General System Level Model
The input variables are described in Table 1 and their respective fuzzifications are presented in
Table 2. Due to paper size limitations only one example of the fuzzified variables for each axis is
shown (Table 2), since the others are similar.
Variables Description
Max_Theta_X/Y/Z
(variables 1,2,3)
For each of the three satellite axes X/Y/Z this represents the maximum
value of the estimated Attitude.
Max_Drift_X/Y/Z
(variables 4,5,6)
For each of the three satellite axes X/Y/Z this represents the maximum
value of the estimated Drift.
Max_Delta_Drift_X/Y/Z
(variables 7,8,9)
For each of the three satellite axes X/Y/Z this represents the (absolute
value of the) difference between the maximum and minimum value of theestimated Drift
Max_Coherency_Index
(variable 10)
The maximum value of the coherency index.
Coherency_Index_Trigger
(variable 11)
The number of times when the coherency index exceeds a given (user
configurable) threshold.
Table 1. Brief description of the input variables
8/9/2019 Mandani
4/8
Table 2. The fuzzified input variables of the model.
In this research 73 rules are used and the output variables and their definition are as follows:
Alarm level Mamdani TSK
Name: Function Type: Parameters: Parameters
(singletons):
Nominal Trapezoidal [0 0 20 30] [12.36]
Non-GyroSoftAlert Trapezoidal [10 20 40 50] [30]
SoftGyroAlert Trapezoidal [30 40 60 70] [50]
GyroAlert Trapezoidal [50 60 80 90] [70]
SeriousGyroAlert Trapezoidal [70 80 100 110] [90]
CriticalGyroAlert Trapezoidal [90 100 120 120] [107.6]
Table 3. Output variables definitions for both models
It should be noted, that we slightly modified the original model outputs (they were almost uniform
functions) to highlight the comparison between both FIS. Further, we used the centers of gravity
of the output membership functions as the TSK output parameters.The input variables are identical in both FIS and the difference is in the consequents of the rules
as can be observed in Table 3. The min operator is used for the implication in both FIS. For the
Mamdani FIS, the aggregation was done using the max operator and for defuzzifier the center of
gravity was used. For the TSK FIS the aggregation was done with the classical weighted average,
using the singletons presented in Table 3.
4. COMPARISON TESTS
To compare the performance of the two types of rule base models, we use three kinds of tests, as
mentioned in the introduction. Details about each test and discussion of results are presented in
the next sub-sections.
8/9/2019 Mandani
5/8
4.1 Processing Time test
In this test a set of randomly generated data was used (increasing the number of input vectors
from 0 to 35000). For each measurement the same set of data was used for both systems.
Throughout the procedure, increasingly larger sets (10 times more for each test) of input data weretested to see how the execution time evolved.
As for the actual measurements, they were taken as the mean of ten measurements obtained for
each set of input data and for each system. Figure 2 presents the results obtained for both systems.
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
0 5000 10000 15000 20000 25000 30000
Number of Input Vectors
ExecutionTime(seconds).
Execution time for Mamdani FIS Execution time for Sugeno FIS
Figure 2. Comparison of process time operation between Mamdani and TSK
On the average, the Mamdani system required 14 times more processing time than the TSK
system. More specifically, on the average, the time taken for the Mamdani system to return one
result was 4.6 x-4
seconds and 3.2 x-5
seconds for the TSK system.
4.2 Robustness to noise
The data used in this test was a set of ten thousand randomly generated input vectors. These
points were changed with randomly generated noise, using a normal distribution with zero mean
and increasing the variance 10 times for each test (Figure 3 shows the variance increases in the xx
axis).
0
20
40
60
80
100
120
0.00000 0.00001 0.00010 0.00100 0.01000 0.10000 1.00000
Variance (Logaritmic scale)
Mean
Squared
Error
Mamdani FIS Sugeno FIS
Figure 3. Comparison of Mamdani with two types of TSK by using sensitive analysis
8/9/2019 Mandani
6/8
Figure 3 shows the comparison for both FIS using the mean squared difference between the
outputs obtained from data corrupted with noise and the results for the data without noise. As can
be observed in Figure 3, the TSK FIS is almost always closer to its noise-free results, meaning
that the noise present in the input data is not affecting the behavior of the system and, hence, the
TSK FIS is quite robust. On the other hand, for very high noise variance the TSK FIS deviates
more than the Mamdani FIS; this means that the data for the system is becoming so different from
its original noise-free values that the system begins producing quite different results. Hence, we
may say that the TSK system is better than the Mamdani FIS with respect to noisy input data.
Further, the TSK FIS is responsive to the fact, that when the noise becomes too high (i.e. when
the input data is drastically changed), the TSK system reacts more strongly and, hence, its
behavior is more realistic.
4.3 Sensitivity analysis
In this test a time series was considered (with 1000 points) when all values for the input variables
are increasing, at a constant rate, from their minimum to their maximum. Figure 4 depicts the
results for both systems.
increase from minimum value to maximumFigure 4. Sensitivity comparison between Mamdani and TSK FIS
As can be observed, Figure 4 shows that there is a decrease in the alarm level, for the TSK
system, between iterations 300 and 500 (area highlighted in gray).
a- MaxTheta b- MaxDrift
Figure 5. Two types of input values between iterations 300 and 500
8/9/2019 Mandani
7/8
Regarding iterations 300 to 500 (approximately), the values for the input variables 1 to 3 and 7 to
11 (see Table 2) are entering a range where the membership degrees for Tolerable and not
Tolerable are close to one another (see for example the variable shown in Fig.5.a highlighted
area). On the other hand, for variables 4 to 6 (Table 2) a high membership value for the term
Tolerable is being reached (see Figure 5.b. highlighted area), as one would intuitively expect,
since the only variables where a clear result is being obtained are the Tolerable, and the other
variables are in a state of limbo between Tolerable and Not Tolerable. The TSK system is
sensitive to this area by decreasing the level of alarm while the Mamdani system is completely
numb and does not respond to these changes. It should also be noted that the TSK FIS has
smoother transitions than the Mamdani FIS.
5. CONCLUSIONS
This research showed that for this case study TSK FIS does not only works better in case of
processing time but also perform better in the other tests, showing that the structure of the TSK
FIS is more robust in the presence of noisy input data (until a certain point, obviously).Furthermore, when we tested the sensitivity of both FIS systems we observe that the TSK FIS is
more sensitive in areas where there is significant imprecision in the input representation, i.e. when
the fuzzy sets overlap.
In summary, we believe that the TSK FIS should be used whenever we have applications with a
single output variable (MISO systems). For MIMO systems (usual in classification problems) the
Mamdani FIS is maybe more appropriate, i.e. it can deal directly with MIMO, while TSK FIS
would require dividing the MIMO into as many MISO systems as the number of output variables;
often a cumbersome and time consuming task. We are presently studying ways to transform
automatically a MISO Mamdani FIS into a TSK FIS and vice versa.
6. ACKNOWLEDGEMENTS
This research was developed in part under the project New Operators for Monitoring and
Diagnostic Intelligent Systems, contract No: 18989/05/NL/MV, of the European Space Agency
(ESA/ESOC). The example used was adapted from the research project "Fuzzy Logic for Mission
Control Processes" financed by European Space Agency, ESA/ESOC No: AO/1-3874/01/D/HK
(2001-2003).
7. REFERENCES
[1] A. Pereira, F. Moura-Pires, R. A. Ribeiro, L. Correia, N. Viana, F. Javier Varas, G.
Mantovani, P.L. Bargellini, R. Perez-Bonilla, A. Donati Fuzzy Expert System for Gyroscope
Fault Detection. 16th
European Simulation Multiconference: Modelling and Simulation 2002,
ISBN 90-77039-07-4 . Darmstadt, Germany June 3-5, 2002.
[2] F. Moura-Pires, R. A. Ribeiro, A. Pereira, F. J. Varas, G. Mantovani, A. Donati Data quality
fuzzy expert system. Proceedings of the 10th Mediterranean Conference on Control and
Automation (MED02), Lisbon, Portugal, July 2002.
[3] E. H. Mamdani and S. Assilian, An experiment in linguistic synthesis with a fuzzy logic
controller,International Journal of Man-Machine Studies, Vol. 7, No.1, 1975, pp. 1-13.
[4] T. Takagi and M. Sugeno, Fuzzy identification of systems and its applications to modeling
and control,IEEE Trans, on Systems, Man and Cybernetics, 15, 1985, pp. 116-132.
[5] J. Mendel, Uncertain rule-based fuzzy inference systems: Introduction and new directions,
Prentice-Hall, 2001.
8/9/2019 Mandani
8/8
[6] D. Tikk, L. T. Kczy, T. D. Gedeon, A survey on universal approximation and its limits in
soft computing techniques. Research Working Paper RWP-IT-012001, School of Information
Technology, Murdoch University, Perth, W.A., 2001.
[7] M. Takcs, Critical Analysis of Various Known Methods for Approximate Reasoning inFuzzy Logic Control. 5th International Symposium of Hungarian Researchers on Computational
Intelligence, Budapest, Hungary 2004.