IMS Project Portfolio - 2015

NSF I/UCRC For Intelligent Maintenance Systems

University of Cincinnati P 1.513.556.3412

PO Box 210072 F 1.513.556.3390

Cincinnati, OH 45221 www.imscenter.net

1

IMS Project Portfolio

Center for Intelligent Maintenance Systems (IMS)

University of Cincinnati

February 2015 Copyright © 2015



PO Box 210072 F 1.513.556.3390


2

Contents

1.1 The Evolution of Maintenance Practices ...................................................................................... 4

1.2 Unmet Needs in Current Maintenance Service Practices ...................................................... 5

1.2.1 Equipment Intelligence .............................................................................................................. 5

1.2.2 Synchronization Intelligence .................................................................................................... 7

1.2.3 Operations Intelligence .............................................................................................................. 7

1.3 The Future of Maintenance Service................................................................................................ 7

1.4 The IMS Center: Methodology and Tools ..................................................................................... 8

1.4.1 IMS Methodology: ‘5S’ Systematic Approach ..................................................................... 9

1.4.2 IMS Tools: Watchdog Agent® ............................................................................................... 10

2.1 Industrial Projects .............................................................................................................................. 13

2.1.1 Project Collaboration with TOYOTA ................................................................................... 13

2.1.2 Project Collaboration with GENERAL MOTORS (G.M.) ............................................... 16

2.1.3 Project Collaboration with HARLEY-DAVIDSON MOTOR COMPANY (HDMC) .. 17

2.1.4 Project Collaboration with KOMATSU ............................................................................... 18

2.1.5 Project Collaboration with CATERPILLAR (Phase 1) .................................................. 19

2.1.6 Project Collaboration with CATERPILLAR (Phase 2) .................................................. 21

2.1.7 Project Collaboration with ALSTOM TRANSPORT ........................................................ 21

2.1.8 Project Collaboration with PARKER-HANNIFIN ............................................................ 26

2.1.9 Project Collaboration with OMRON .................................................................................... 27

2.1.10 Project Collaboration with GENERAL ELECTRIC (GE) ............................................. 29

2.1.11 Hong Kong’s Airport Chiller Health Monitoring .......................................................... 31

2.1.12 Project Collaboration with TECHSOLVE ........................................................................ 33

2.1.13 Intelligent Prognostic Tools for Smart Machine Platform ....................................... 35

2.1.14 Project Collaboration with SIEMENS ............................................................................... 40

2.1.15 Project Collaboration with Hiwin, Taiwan .................................................................... 41

2.1.16 Project with INSTITUTE FOR INFORMATION INDUSTRY (III), TAIWAN ......... 44

2.1.17 Project Collaboration with FMTC ...................................................................................... 49

2.1.18 Project Collaboration with Korea Aerospace University (KAU) ........................... 50

2.1.19 Project Collaboration with Cosen ..................................................................................... 50

2.1.20 Project Collaboration with Daniel Drake Center for Post-Acute Care and Beechwood Home ..................................................................................................................................... 53



PO Box 210072 F 1.513.556.3390


3

2.2 Fundamental Research Projects ................................................................................................... 58

2.2.1 A Systematic Methodology for Data Validation and Verification for Prognostics Applications ................................................................................................................................................ 58

2.2.2 Hybrid Modeling of Equipment Efficiency and Health for Advanced Servo-drive Presses .......................................................................................................................................................... 60

2.2.3 Battery Prognostics ................................................................................................................... 63

2.2.4 Remaining Useful Life Prediction of In-vehicle Battery Pack ................................... 66

2.2.5 Virtual Bearing ............................................................................................................................ 68

2.2.6 Couple model ............................................................................................................................... 69

2.3 PHM Society Data Challenges ........................................................................................................ 72

2.3.1 [PHM 2008] A Similarity-based Prognostics Approach for Remaining Useful Life Estimation of Engineered Systems .................................................................................................... 72

2.3.2 [PHM 2009] Gearbox Health Assessment and Fault Classification ........................ 74

2.3.3 [PHM 2010] Remaining Useful Life Estimation of High-speed CNC Milling Machine Cutters ........................................................................................................................................ 76

2.3.4 [PHM 2011] Fault Detection of Anemometers ............................................................... 79

2.3.5 (2014) An Effective Predictive Maintenance Approach based on Historical Maintenance Data using a Probabilistic Risk Assessment ........................................................ 81

2.4 Biography: Professor Jay Lee ......................................................................................................... 85



PO Box 210072 F 1.513.556.3390


4

1. Introduction

1.1 The Evolution of Maintenance Practices

The traditional approach for maintaining the industrial machines and processes has been one that is purely reactive. Mitigating procedures are performed to the equipment once it breaks down. This, as most companies have found out, is a maintenance paradigm with many disadvantages. When a machine is run to failure, i.e. until one of its parts fail making it inoperable, produced parts may not meet quality standards, or the degradation of the other components might have just been accelerated unknowingly, or worst, the cost of restoring the machine to an acceptable and sustained operational state might be more expensive than a purchasing a replacement unit. There is also a requirement to store numerous spare parts. Finally, since a machine can break down at any time, maintenance and production operations are not scheduled efficiently, thus, a huge maintenance crew is needed to achieve high production throughput rates.

An improved strategy that companies employ is preventive maintenance. Many of the machine manufacturers and suppliers provide component lifetime and maintenance guidelines to their customers. With this method, maintenance activities can be performed so as not to interfere with production schedule. There is no need to store a large inventory of spare parts because they can be ordered at regular intervals just before the parts need to be replaced. The primary advantage of using preventive maintenance is a high overall equipment efficiency (OEE) score, but, it does so at a very high cost. To be able to capture the majority of impending component failures, maintenance personnel must replace or condition the parts at closer time intervals, which mean that spare parts must be purchased more frequently.

Nowadays, many manufacturing companies are adopting condition-based maintenance or CBM. Predictive Maintenance (PdM) is a right-on-time maintenance strategy. It is based on the failure limit policy in which maintenance is performed only when the failure rate, or other reliability indices, of a unit reaches a predetermined level. This maintenance strategy has been implemented as Condition Based Maintenance (CBM) in most production systems, where certain performance indices are periodically (Barbera et al. 1996; Chen et al. 2002) or continuously monitored (Marseguerra et al. 2002). Whenever an index value crosses its predefined threshold, maintenance actions are performed to restore the machine to its original state, or to a state where the changed value is at a satisfactory level in comparison to the threshold.

Predictive maintenance (PdM) can be best described as a process that requires both technology and human skills, while using a combination of all available diagnostic and performance data, maintenance history, operator logs and design data to make timely decisions about maintenance requirements of major/critical equipment. It is this integration of various data, information and processes that leads to the success of a PdM program. It analyzes the trend of measured physical parameters against known engineering limits for the purpose of detecting, analyzing and correcting a problem before a failure occurs. A maintenance plan is made based on the prediction results derived from condition-based monitoring. This method can cost more up front than PM because of the additional monitoring hardware and software investment, cost of manning, tooling, and education that is required to establish a PdM program. However, it provides a basis for failure diagnostics and maintenance operations, and offers



PO Box 210072 F 1.513.556.3390


5

increased equipment reliability and a sufficient advance in information to improve planning, thereby reducing unexpected downtime and operating costs. This approach takes into account the current state or condition of the component, thereby reducing the number of unscheduled breakdowns by monitoring the equipment condition to predict failures so that remedial fixes can be performed. Using built-in or add-on sensors, signal measurements are extracted to describe the component’s current performance. Then, this information is compared with previously collected performance history to determine whether a deviation has occurred to suggest that the equipment is at fault or showing symptoms that can lead to an impending failure. The apparent shortcoming of CBM is that it simply detects that the equipment is “at fault” or “nearing a fault” condition. It does not elaborate on the type of failure the equipment has experienced and it does not identify what component is faulty. The information from a CBM system does not provide much insight to the maintenance personnel on what to fix in that machine. Predictive maintenance has been developed as an improvement to CBM so as to infer within a horizon time when the equipment will fail. Nonetheless, it inherited the same problems as the CBM method in pinpointing the type of fault and the critical component at fault.

1.2 Unmet Needs in Current Maintenance Service Practices

After reviewing the evolution of the different maintenance paradigms, there exist key issues that need to be addressed in order to achieve the maximum availability and efficiency from important company assets. These unmet needs are summarized in Figure 1 and are explained in detail in the succeeding sections.

1.2.1 Equipment Intelligence

There is indeed a need to migrate towards another maintenance paradigm – intelligent prognostics. It transcends being merely a predictive maintenance system by exactly identifying which component of a machine is likely to fail, and when and highlight the event to trigger maintenance service and order spare parts. [2] A key principle in intelligent maintenance is the assessment and prediction of the performance degradation of a process, machine or service. [3] By extracting critical features from signal measurements, performance degradation can be used to predict unacceptable equipment performance before it occurs. With the use of enabling and transformation technologies, plant maintenance can be shifted from the “fail-and-fix” tradition into the “predict-and-prevent” principle as shown in Figure 2.



PO Box 210072 F 1.513.556.3390


6

Figure 1. Unmet Needs in Maintenance Services

Figure 2. Maintenance Service Paradigm Shift from Fail-and-Fix to Predict-and-Prevent

By utilizing sensors and embedded systems, the intelligent prognostics approach shall allow the critical asset to be able to automatically trigger itself to acquire appropriate health indicators from its critical machine parts when needed, assess the features from the temporal health data, diagnose itself to determine its current state of machine performance, infer whether significant performance degradation has been detected, identify the imminent fault it is going to experience and predict when failure is going to occur and automatically call for maintenance crew for conditioning. These capabilities just mentioned are essential to the future of maintenance service where machines are capable of self-assessment or self-diagnostics and ultimately, self-maintenance and even self-healing. Self-healing can be achieved when the machine is supplied with appropriate control logic to follow whenever its components have reached specific degradation levels such as shutting down secondary subsystems that are not critical to its basic functionalities to reduce load and system power consumption.



PO Box 210072 F 1.513.556.3390


7

1.2.2 Synchronization Intelligence

Currently most of the machine data, if available, is either hidden or kept in the machine. Thus, assets in a manufacturing plant become islands of health indicators and are not being accessed and analyzed for health information. If ever the data undergoes a certain degree of processing, the information that is extracted unfortunately remains trapped in the machine. There is a recognized barrier between the production floor and the management level that inhibits critical machine and process health information to be transmitted so that the latter can make more informed and timely decisions on the machines. What is needed is an infrastructure that enables the health information to be seamlessly transmitted to a higher manufacturing level for decision making, or to other similar machines for peer-to-peer equipment performance comparison.

When machine performance information is sent up the production business unit, comprehensive and real-time decisions can be made on the manufacturing assets such as production scheduling (assigning appropriate order sizes to high performing machines), maintenance scheduling (with a sufficient horizon time, conditioning or mitigating procedures can be performed while not interfering with production runs) and even machine-part matching (assigning the “best” machine that will meet quality requirements of a particular produced part).

1.2.3 Operations Intelligence

When the company assets are now capable of self-diagnostics and self-maintenance, and the health information (both machine and process) can be shared very easily most especially to the different business functions, then there lies an opportunity to achieve intelligent operations in the manufacturing facility.

With the machine health information, one is enabled to predict and prioritize maintenance activities, and when coupled with the process health information, production can be optimized either to achieve desired volume orders and/or part quality requirements. Thus, the company achieves intelligent operations in the process because they are now capable of producing quality parts while running at a near-zero unexpected downtime run.

The missing component to achieving such a goal is an optimized decision making logic modeled on plant operations that integrates production and maintenance activities without interfering with each other.

1.3 The Future of Maintenance Service

In summary, the key issues of the current manufacturing maintenance practices have been defined: (1) equipment intelligence through self-maintenance – to be able to extract machine and process health and eventually predict machine performance degradation; (2) synchronization intelligence – refers to the seamless communication between the production floor and management unit so that timely and informed decisions can be made based on machine and process health information; and (3) operations intelligence – the ability to leverage on self-maintenance and intelligent synchronization to efficiently schedule production and maintenance to achieve a near-zero downtime manufacturing operations.

The Center for Intelligent Maintenance Systems (IMS) has been in the pursuit to address these key issues by performing relevant research in order to develop the enabling technologies for the different aspects of self-maintenance, intelligent synchronization and intelligent operations.



PO Box 210072 F 1.513.556.3390


8

Figure 3 summarizes the vision of the IMS Center while providing a directed plan on how to solve these maintenance opportunities.

Figure 3. IMS Center Vision of Intelligent Maintenance

From the IMS vision, degradation information from the critical assets can be used to predict their performance to enable near-zero downtime performance. The asset degradation data can also be fed back to the design stage of the equipment for assessing its reusability and predicting its useful life after disassembly and reuse.

To address the key issues mentioned earlier, the IMS Center has recognized that the following technologies need to be developed:

Processing system that extracts machine data, and convert this voluminous data space

into a simple but comprehensive health information;

Communication infrastructure that can automatically relay health information to

appropriate business functions within the company; and

Decision-making system that uses the health information to make informed business

decisions on the company assets.

1.4 The IMS Center: Methodology and Tools

The IMS Center is an Industry/University Research Cooperative Center (I/UCRC) for Intelligent Maintenance Systems that has been founded to serve as a center of excellence for the creation and dissemination of a systematic body of knowledge in intelligent e-maintenance systems and ultimately to impact next-generation product, manufacturing, and service systems. The IMS Center serves as a catalyst as well as enabler to assist company members to transform their



PO Box 210072 F 1.513.556.3390


9

operation strategies from today’s “Fail-to-Fix/Fly-to-Fix (FAF)” to “Predict-and-Prevent (PAP)” performance. The IMS Center brings value to its members by validating high-impact emerging technologies as well as by harnessing business alliances through collaborative test-beds.

Through this end, the IMS Center shall enable products and systems to achieve and sustain near-zero breakdown performance, and ultimately transform maintenance data to useful information for improved closed-loop product life cycle design and asset management. The IMS Center is focused on frontier technologies in embedded and remote monitoring, prognostics technologies, and intelligent decision support tools and has coined the trademarked Watchdog Agent prognostics tools and Device-to-Business (D2B) Infotronics platform for e-maintenance systems.

1.4.1 IMS Methodology: ‘5S’ Systematic Approach

Addressing future maintenance services necessitates a systematic “5S” approach (see Figure 4). This approach was devised by IMS Center in order to develop and research all aspects of future maintenance infrastructures. This systematic approach consists of five key elements:

Streamline: this encompasses techniques for sorting, prioritizing, and classifying data into

more feature-based health clusters. This may also include reducing large data sets (from

both maintenance history and on-line data DAQ) to smaller dimensions, leading to

correlation of the relevant data to feature maps for better data representation.

Smart Process: using the right Watchdog Agent® tool for the right application. This requires

techniques for selecting appropriate prognostics tools based on application conditions,

criticality of each conditions for machine health, and system requirements.

Synchronize: converting component data to component degradation information at the local

level and further predicting trends of health using a visualized radar chart for decision-ready

information. Maintenance data is transformed to health information and to an automated

action (i.e. order parts, schedule maintenance based on criticality of component or

machine). This process ensures a key characteristic of “Only Handle Information Once”

(OHIO).

Standardize: creation of a standardized information structure for equipment condition data

and health information so that it is compatible with higher-level business systems and

enables the information to be embedded in business ERP and asset management systems.

The goal here is to keep the process as a standard approach for day-to-day practices.

Sustain: utilizing the transformed data for information-level decision making. System

information is then shared amongst all stages of product and business life cycle systems:

product design, manufacturing, maintenance, service logistics, and others in order to realize

a closed-loop product life-cycle design and continuous improvement of product-level quality

and business-level performance.



PO Box 210072 F 1.513.556.3390


10

Figure 4. IMS 5S Systematic Methodology

1.4.2 IMS Tools: Watchdog Agent®

Since 2000, the IMS Center has been spearheading the development of a processing system called the Watchdog Agent® in a bid to address the first issue. Simply put, the Watchdog Agent® is an enabling technology that shall allow for a successful implementation of intelligent maintenance. The toolbox is a collection of algorithms that can be used to assess and predict the performance of a process or equipment based on input from sensors, historical data and operating conditions. Referring to Figure 5, the algorithms can be classified into four major categories: signal processing and feature extraction, quantitative health assessment, performance prediction and condition health diagnosis. Performance-related information can be further extracted via multiple sensor inputs through signal processing, feature extraction and sensor fusion techniques. The historical behavior of process signatures can be utilized to predict future behavior and thus enable forecasting of the process’s or machine’s performance. Proactive maintenance can therefore be facilitated through the prediction of potential failures before they occur.

By leveraging on current technologies, the Watchdog Agent® can either exist in Matlab or LabVIEW platform although development is underway to seat the Watchdog Agent® in PLC controller and C-based platforms. This has been done to address the diversity of platform that are currently being used in most manufacturing facilities.



PO Box 210072 F 1.513.556.3390


11

Figure 5. Software Architecture and Visualization Tools of Watchdog Agent® Toolbox



PO Box 210072 F 1.513.556.3390


12

2. IMS Research and Project Portfolio

This section briefly summarizes several projects and test-beds where the Watchdog Agent® has been used. The general methodology of IMS intelligent prognostics is still being applied; only the algorithms are selected and reconfigured depending on the monitoring task being implemented.

The Center for Intelligent Maintenance Systems (IMS) is working with its research partners to file joint patents and technology (IP) transfer for industrial applications. Examples of such efforts are IMS activities with Parker-Hannifin, Techsolve, and Siemens.

INDUSTRY COMPANY PROJECT

Automotive

Toyota Surge map modeling for air compressors

General Motors Prognostics of vehicle components

Harley Davidson Spindle bearing monitoring

Heavy Machinery

Komatsu Improving the diagnostic and prognostic monitoring

capabilities for Komatsu’s truck diesel engines

Caterpillar Machine Tool health monitoring

ALSTOM Transport Feasibility study of PHM applications

Process

Proctor and Gamble Quality-centric process health management

Parker Hydraulic hose prognostics

Omron Precision energy management systems

General Electric Intelligent Prognostic System for Machine Health

Monitoring

Hong Kong Electrical and

Mechanical Services

Department (EMSD)

Hong Kong Airport’s Chiller Health Monitoring

Consulting

Techsolve Smart machine platform initiative (SMPI)

Techsolve Intelligent Prognostic Tools for Smart Machine

Platform

Siemens Reconfigurable plug-n-prognose Watchdog Agent®

There have also been some researches apart from the industry including the PHM data challenges and NSF fundamental researches. They are summarized in the table below:



PO Box 210072 F 1.513.556.3390


13

Research Title

Fundamental

Researches

A Systematic Methodology for

Data Validation and Verification

for Prognostics Applications

In collaboration with IMS

center at University of

Michigan and Missouri

University of Science and

Technology

Hybrid modeling of equipment

efficiency and health for

advanced servo-drive presses

In collaboration with Center

for Precision Forming at

Ohio State University

Industrial Energy Monitoring In collaboration with Omron

Battery Prognostics

RUL prediction of the in-vehicle

battery pack

In collaboration with

Shanghai Jiao Tong

University

Virtual Bearing

PHM Society Data

Challenges

[PHM 2008] A similarity-based

prognostics approach for

Remaining Useful Life estimation

of engineered systems

1st and 3

rd Rank

[PHM 2009] Gearbox Health

Assessment and Fault

Classification

1st and 2

nd Ranks

[PHM 2010] RUL estimation of

high-speed CNC milling machine

cutters

3rd

Ranks

[PHM 2011] Anemometer fault

detection 1

st and 3

rd Ranks

2.1 Industrial Projects

2.1.1 Project Collaboration with TOYOTA

a) General Background:

This project is conducted between the Center for IMS and Toyota Motors Manufacturing, Kentucky, Inc. (TMMK), the objective of which is to improve the accuracy and efficiency of surge avoidance control for centrifugal air compressors. The main achievements of the project work is



PO Box 210072 F 1.513.556.3390


14

to use data-driven tools to model an improved surge map, and provide positive feedback control of inlet guide vane (IGV) to reduce the rate of surge events in the manufacturing facility of TMMK.

b) Project Duration:

February 2005 – December 2007

c) Project Objectives:

The objective of this research was to develop a data-driven modeling based approach for obtaining surge maps for air compressors under variable operating conditions so as to achieve surge prevention with high efficiency.

d) Summary of Technical Approach:

Surge refers to a phenomenon of large oscillating reverse flow when compressor is operated under low flow rate and high pressure. Surge leads to costly damage of the compressor and downtime of the system powered by the compressor. To prevent the damage, a surge map is widely used where a surge line is defined to separate surge and not-surge operation (Figure 6).

Figure 6. Surge Maps for Compressors

However, there is a challenge that exact location of the surge line cannot be accurately located due to the fact that large number of variables related to compressor operation, including ambient air conditions can affect the occurrence of surge. Hence the operators would have to operate the compressor far from the surge line to avoid the possibility of experiencing surge, which on the other hand would decrease compressor efficiency. In order to operate the compressor at optimal efficiency while preventing surge, a data-driven method is used to develop an improved surge map model for actual operation in the specific environment.



PO Box 210072 F 1.513.556.3390


15

Figure 7. Data-driven Surge Map Modeling

Figure 7 shows the flowchart of the data-driven approach.

1) Multivariate measurements are synchronously sampled under known surge or not-

surge conditions. Variables include pressure readings, flow rate of airflow, electric

current, IGV feedback, IGV command, humidity, temperature readings etc.

2) Principal Component Analysis (PCA) is employed to reduce the dimension of the

dataset acquired, by projecting the dataset to a new coordination system while

retaining as much variance of the original dataset as possible.

3) Support Vector Machine (SVM) is used to train a classification model based on the

transformed data, or principal components. SVM finds the separating planes that

maximize the distance between the two classes of surge & not-surge, and therefore

can potentially minimize the misclassification.

During around 80 days of data acquisition, 25 surge and 22 not-surge data points were collected and 15 from each class are used for training the separation plane. For the training dataset, there is no misclassification hence there is no unpredicted surge and false alarms. The rest data points were used to test model performance and there were 6 false alarms among the 22 not-surge data. The model was further developed and used for open-loop surge avoidance control.

e) Project Deliverables and Results:

Eventually as deliverables of the project, a validated IMS tool is ported into existing compressor controller. The compressor manufacturer that supplies for TMMK was suggested to provide such capability of on-line surge detection and control optimization. Existing machines were retrofitted with surge detection technique, which saves the facility around 50,000 US dollars per year as opposed to actual project budget 100,000. The project is recognized as a success story and promoted constantly at TMMK facility.



PO Box 210072 F 1.513.556.3390


16

2.1.2 Project Collaboration with GENERAL MOTORS (G.M.)


This project is collaboration between IMS center and General Motors Technical Center. The project will aim at implementing IMS tools and methodology for prognostics of vehicle components. The project is divided into several phases; the first phase of the project focused on prognostics of wheel speed sensors of the Anti-lock Braking System (ABS); the second phase of the project is currently dealing with prognostics of the alternator/starter system; later phases of the project will focus on on-board prognostics.


Phase 1: July 2007 – July 2008 (1 Year)

Phase 2: September 2008 – January 2009 (4 months)


Objective 1: Design a prognostic model to assess the health of the wheel speed sensor and differentiate sensor faults from system faults.

Objective 2: Design a prognostic model to assess the health of the alternator/starter system.

Objective 3: Study the feasibility of using IMS tools and methodology to perform on-board prognostics for vehicle components.


For phase I of the project, a thorough study of the existing techniques for sensor diagnostics/prognostics was conducted through a literature review and a patent search. A prognostic decision model based on the Watchdog Agent ® was built. The model consists of two steps: off-line training and on-line testing.

In the step of off-line training, the basic task is to establish the data model for the normal characteristics based on the features extracted from the direct sensor readings, and train the artificial intelligence algorithms off-line. For this project two algorithms from the Watchdog Agent ® were used: logistic regression (LR) and statistical pattern recognition (SPR). Data was initially collected directly from the vehicle’s sensors; then a test-rig was assembled at the IMS center that allowed inducing faults to sensors and collecting data. Features extracted from the active wheel speed sensor data are: mean pulse duty factor, mean upper level amplitude, and mean lower level amplitude. Different features can reflect different physical characteristic of monitored sensor. Pulse duty factor can reflect the wear or broken of toothed ring, amplitude related features can reflect the magnetic field change of the sensor. Therefore, changes of certain feature can be used to indicate the potential faulty element.



PO Box 210072 F 1.513.556.3390


17

In the step of on-line testing, the real-time sensor readings are captured and the corresponding features can be extracted, which are further input into the trained model. A CV value can be obtained to show the current health status. The features are also analyzed and visualized through SOM. Finally, a comprehensive analysis decision is given, which include the CV value, the possible physical reason if CV is abruptly changed and corresponding solutions.

For phase II of the project, a test-rig is currently being built at the IMS center to run the alternator and collect data from both a healthy and faulty alternator. After that a prognostic model will be developed at the IMS center for the alternator/starter system.

2.1.3 Project Collaboration with HARLEY-DAVIDSON MOTOR COMPANY

(HDMC)


The integration of the Watchdog Agent® prognostics platform and wireless sensor network successfully realized real-time multiple networked machines health monitoring and prognostics. It would be directly helping production become more cost effective at Harley-Davidson motor company as it says in the article “Harley-Davidson: Born to be … predictable” in Lean Tools for Maintenance & Reliability magazine.


June 2006 – August 2006 (2 Months), June 2007 – August 2007 (2 Months)

c) Objectives:

Automate the data acquisition without interfering with the machine operation.

Save the investment of health monitoring and prognosis - upgrade the machine health monitoring capabilities by using a single platform to monitor multiple machines with multi-speed and various spindle loads.

Seamless integration of wireless vibration accelerometer with the Watchdog Agent® platform.

Reduce the data analysis time by providing autonomous data processing capabilities with logistic regression and statistical pattern recognition algorithms.

Enhance the trend prediction method by providing autoregressive moving average model (ARMA) and neural network algorithms.


Generate Feature Vector: The feature vector is a signature describing a state of bearing health, which is formed from features that are considered to be correlated to those that indicate machine health condition. Features are extracted by Fast Fourier Transform (FFT).

Evaluate Bearing Health: Patterns of the feature vectors that describe up-to-now states of bearing health are grouped into a health map, using a clustering algorithm such as the method



PO Box 210072 F 1.513.556.3390


18

of self-organizing map, k-means algorithm, neural networks-based algorithm, and others, depending on available information about patterns of the feature vectors.

Trend Prediction: Autonomous algorithms (ARMA and Neural Network) are used to predict the trend of the extracted feature or the trend of the health assessment results (confidence values).

The uniqueness of the technique is the Watchdog Agent platform that provides the health information of the bearing performance instead of the raw signal data obtained from the sensors. Utilizing the current signal as a triggering mechanism, the system is able to monitor the vibration signals within any machining process of the machine without interaction with the controller. Based on the vibration signals, data processing tools are used to convert data into health information for the bearing performance. Hence, the information is presented in a radar chart for visualization. This platform is also designed for integration in the enterprise asset management systems.

2.1.4 Project Collaboration with KOMATSU


This project was quite different in format, in that a Komatsu engineer spent 1 year at the Center for Intelligent Maintenance Systems at the University of Cincinnati to understand the algorithms and data processing methods that were applicable to his application. Based on the guidance from researchers at the center, the Komatsu engineer would then apply those techniques to the Komatsu heavy truck diesel engine data collected in the field.

b) Project Objectives:

Study the feasibility of using IMS tools and methodology to improve the diagnostic and prognostic monitoring capabilities for the diesel engines used on Komatsu heavy duty trucks.

Development of a decision aid tool that can be provided to the engine maintenance technician.

c) Summary of Technical Approach:

This particular application was for a heavy-duty equipment vehicle used in mining and construction. The remote prognostics and monitoring system focused on assessing and predicting the health of the diesel engine component. For this remote monitoring application, the previously developed architecture for data acquisition and data storage consisted of sending a daily data set of parameters from the diesel engine to the remote location. The parameters included pressures, fuel flow rate, temperature, and rotational speed of the engine. These parameters were taken at key operating points for the engine, such as at idle engine speed or at maximum exhaust gas temperature. The previously developed architecture was missing the necessary algorithms to process the data and assess the current health of the engine, determine the root cause of the anomalous behavior, as well as predict the remaining life of the



PO Box 210072 F 1.513.556.3390


19

diesel engine. The heavy-duty equipment manufacturer in collaboration with the Center for Intelligent Maintenance Systems (IMS) developed a systematic approach utilizing several algorithms from the suite of algorithms in the Watchdog Agent® toolbox to convert the diesel engine data into health information.

The data preprocessing step consisted of using the Huber method for outlier removal, as well as the use of an auto-regressing moving average approach to predict a time series value a few steps ahead to replace missing values. The missing values could be due to an error in the transmission of the data to the remote location or from an outlier removal preprocessing step. After preprocessing the data, the next step was to develop a methodology to classify the different engine patterns in the data to particular engine related problems. The use of a Bayesian Belief Network (BBN) classification technique used the manufacturer’s experience on engine related problems along with the pattern history of the data to build the model. This classification model was able to interpret the anomalous engine behavior in the data and identify the root cause of the problem at the early stage of degradation.

The last remaining step is the remaining life prediction, and this used a fuzzy logic based algorithm. The fuzzy membership functions were based on engineering experience as well as features extracted from the data patterns; this hybrid approach accounts for the uncertainty in the data and combines data driven and expert knowledge for a more robust approach.

d) Project Deliverables and Results:

An overall visualization of the final output is shown in Figure 8, highlighting the decision aid that can be provided to the maintenance technician.

Figure 8. Example of Engine Monitoring Interface

2.1.5 Project Collaboration with CATERPILLAR (Phase 1)


The project is subcontracted to the IMS Center as part of Caterpillar’s project called Manufacturing Asset Capability Information Tools (MACIT). IMS provides consultant and

Wear

Cracking

Cable disc.

Sensor fails

Injector

Wear

Cracking

Sensor fails

D: Detectability C: cost

P: Probability S: Severity

S Prognostic

Machines

Effect on the

system

Component D PSystem ref. Engine ref. Failure Mode

Machine B

Engine AXMachine A

CCauses

2FM 1

FM2

Chute de la

cabine

Transmis BY

Cylinder

Inlet Valve R

Inlet Valve L

Injector

…

4 0.9

DYNAMIC BBN & FL Based FMERA

H Change

engine

Déraillement

du câble

Chute de la

cabine (la

chute d'une

Shaft

Gears

Bearing

1 M Change

Bearing

0.6

BBN output Decison Aid



PO Box 210072 F 1.513.556.3390


20

technical assistance to Caterpillar. In addition, two PhD students from the IMS Center were hired as internship by Caterpillar in 2007 and 2008 respectively to help Caterpillar develop their own technology for machine tool health monitoring.


May 21, 2007 - December 14, 2007 (IMS Researcher Internship)

c) Objectives:

Develop an autonomous, online machine tool health monitoring system on a CNC lath test-bed with emphasis on spindle bearing health, tool wear and tool crashes.

Find the appropriate machine tool health monitoring practices that have good generality in principle and are quickly duplicable for a variety of machine tools.


The work starts with the common approach for spindle bearing health and tool wear monitoring based on the previous experience of IMS Center on machine tool prognosis. This approach implements the data-driven prognostic methodology developed at the IMS center to machine tool health monitoring, which includes four key steps: data acquisition, signal processing and feature extraction, health assessment, presentation and reporting. Spindle bearing vibrations, spindle motor current and feeding motor current are instrumented, whereas controller-dependent communications and signals are avoided in order to maintain the generality of the approach. Several algorithms from the Watchdog Agent® toolbox are selected and applied. Since autonomy is required for the health monitoring system, another key step, called process alignment and identification, is inserted immediately after data acquisition; the corresponding technology is developed and validated during the work. This technology enables the health monitoring system to function well when the machine switches among many jobs frequently, i.e. in manual or small-batch production.


A machine tool health monitoring system is built on the test-bed machine.

The system can run online 24 hours per day and 7 days per week, automatically collecting

data, doing analysis and displaying the machine health conditions for spindle bearing and

tools.

The system is also capable to manage the raw data and health information in order for

engineers to track the history and find out reasons of critical events, e.g. spindle crashes.

Several signal streamlining algorithms for process alignment and identification are newly

developed, enabling the whole system to be quickly ported to other type of machines.



PO Box 210072 F 1.513.556.3390


21

2.1.6 Project Collaboration with CATERPILLAR (Phase 2)


The project is subcontracted to the IMS Center as part of Caterpillar’s project called Manufacturing Asset Capability Information Tools (MACIT). IMS provides consultancy and technical assistance to Caterpillar. In addition, two PhD students from the IMS Center were hired as internship by Caterpillar in 2007 and 2008 respectively to help Caterpillar develop their own technology for machine tool health monitoring.


May 19, 2008 - August 9, 2008 (IMS Researcher Internship)

c) Objectives:

Validate and automate the machine tool health monitoring system on the CNC lathe test-bed with emphasis on spindle bearing health, tool wear and tool crashes.

Replicate the prognostic system by developing a machine tool health monitoring system on a grinding test-bed with emphasis on the grinding wheel and coolant condition.


With the system developed during the first internship, the algorithms are validated by correlating the features with the actual machining processes. Features are extracted on appropriate segments of signals that correspond to steady-state (for spindle health) and dynamic state like in a cutting condition (for tool wear). Furthermore, feature selection for tool wear detection is aided by knowledge of the known tool life.

With regards to the grinder test-bed, an online data acquisition system has been developed. Current work is geared towards extracting features from spindle and axis power and current signals to detect grinding wheel wear.


Improved machine health monitoring for the lathe test-bed by increasing process identification accuracy. More relevant features are selected that are more responsive to tool wear.

Grinding machine test-bed has been instrumented. An online data acquisition system is developed.

2.1.7 Project Collaboration with ALSTOM TRANSPORT

IMS has had two training session for ALSTOM Transport personnel during summer 2011; one for ALSTOM R&D engineers of Train Life Services (TLS) in Barcelona, Spain and one for ALSTOM R&D engineers in France. The duration of the training sessions was 6 days. The



PO Box 210072 F 1.513.556.3390


22

training agenda was designed to fit the interests and needs of each group. They were provided with the summary of each PHM method along with sample examples.

There are currently two feasibility studies being conducted by IMS center in collaboration with ALSTOM Transport.

1) Degradation Monitoring of the Electro-Mechanical Mechanism in Point

Machines


Point machines are used for operating the railway turnouts and are considered as critical track elements in railway assets. Failure in the point machine’s mechanism causes delays, increases railway operating costs and more importantly, causes train accidents. Hence the condition monitoring and health management of point machines have become a main areas of interest. This research focuses on establishing a strategy and technical architecture for prognostics and health management (PHM) of the electro-mechanical point machines. This study has been conducted on the P80 electro-mechanical point machine data acquired from eight machines in various locations throughout Italy. Time-stamped data was acquired for point machines current and voltage signals.


August 2011 – January 2012

c) Project Objectives

The objective of this research is to explore the feasibility of applying a PHM strategy and technical approach for the condition monitoring of the point machines. Various feature extraction techniques have been applied to the data. Then, the Principle Component Analysis (PCA) was applied to the features to assess the health of the machines. There is currently lack of information about the current and past conditions of the machines and the exact timing of each participating movement of the mechanism. Despite this, the results obtained show degradation of the machines and demonstrate the applicability of aforementioned PHM technique for fault diagnostics and prognostics of point machines.


The technical approach can be summarized as:

Pre-processing and filtering the raw data

Data segmentation (four segments in this case) as shown in Figure 9

Statistical feature extraction and selection

Health model development using Principal Component Analysis (PCA), as summarized in

Figure 10



PO Box 210072 F 1.513.556.3390


23

Figure 9. The segmentation of a typical current signal for one operation

Figure 10. Feeding the selected features into PCA and obtaining the health value


The machines work in two different directions; normal and reverse. For machine 1, there have been three maintenance events for the operation in reverse direction and none for normal direction. As it can be seen in Figure 11, the health value has an increasing trend in reverse direction and is almost constant in normal direction which clearly matches with what was observed in the maintenance events log.



PO Box 210072 F 1.513.556.3390


24

Figure 11. The obtained health value for machine 1; the green arrows show the times at which maintenance events occurred.

2) Degradation Monitoring of the Stator Winding Insulation of a 3-Phase

Asynchronous Induction Motor


A locomotive is a railway vehicle that moves/pulls train units along the tracks. Locomotives have several critical components that are liable to degradation over time, including the bogie (underlying chassis carrying the wheels) and other internal components such as the motors. Winding insulation degradation is a key concern in the health management of locomotive induction motors. This application focuses on establishing a strategy and technical architecture for prognostics and health management of winding insulation in the 3-phase asynchronous induction motors.


August 2011 – January 2012


The objective of this research is to explore the feasibility of applying a prognostics and health management strategy and technical approach for the condition monitoring of the stator winding insulation of the induction motors with its specific application in the traction motors of high speed trains manufactured by ALSTOM Transport.



PO Box 210072 F 1.513.556.3390


25


The technical approach can be summarized as below:

Data segmentation (as shown in left figure of Figure 12)

Data normalization

Statistical feature extraction and selection. The extracted features include:

o Peak distance

o Shape f actor

o Crest factor

o Impulse factor

o The standard deviation of each feature extracted from the three currents (three

phases)

Some of the features are plotted in right figure in Figure 12.

Figure 12. (Left) The segmentation of the current signal (blue line is the current for one phase raw data, and the pink line is the summation of the current for three phases); (Right) Four of the extracted features from the current signals. As it can be seen, the trends are almost constant.

No trends can be observed in the features as the motor under study is running in healthy condition. A dataset collected from a fleet of 14 locomotives will be provided by ALSTOM which will be more suitable for testing different methods and designing a PHM strategy for monitoring the health of the traction motors.

3) Future Collaborations

Accelerated life testing of Electronic Module Reliability (EMR)

A project with ALSTOM Transport in France

1 2 3 4 5 6



PO Box 210072 F 1.513.556.3390


26

Health monitoring of traction motors for a fleet of 14 locomotives

A project with ALSTOM Transport in Spain

Bogie’s axle bearing health monitoring

A project with ALSTOM Transport in France

2.1.8 Project Collaboration with PARKER-HANNIFIN


Hydraulic power systems are extensively used in many applications. Hose is the “artery” that keeps equipment running. Consequences of hose failure are serious. It not only causes equipment downtime, but also safety issues. Current maintenance scheme is mainly based on preventive or Fail-and-Fix (FAF). In another word, hydraulic hose are replace by schedule or when already failed. A higher level of maintenance, Predict-and-Prevent (PAP) is needed to achieve near-zero down maintenance, which in turn will increase equipment up time and safety.


Sep 1, 2007 - Feb 28, 2009


Figure 13.Smart Hose Vision (Patent)

Figure 13 provides an overview of the Parker Hannifin smart hose vision. The IMS center and the Nano lab at UC work together toward a smart hose solution by comprising smart sensor technology and Watchdog® Agent. The smart sensors are able to perform cycle counting, resistance and strain measurements. It is also able to harvest energy by itself, communicate wirelessly, and is capable of working under harsh environment. A more advanced smart sensor should be able to detect failure for threadlike article. The Nano lab at UC is working avidly to realize such sensors, such as carbon nanotube thread (CNT). Data from sensors is digested in the Watchdog® Agent for Degradation monitoring, fault detection and classification, and maintenance scheduling. A project is formulated to achieve this vision.

Hose pulse and bending

Smart Sensory Technology

• Cycle counting

• Energy harvesting

• Working under harsh environment

• Failure detection for threadlike article

Fixture

Pulling

Forces

0

0.2

0.4

0.6

0.8

1

IMS

Watchdog

Agent®

Toolbox

Health

Information

Watchdog Agent®

• Degradation monitoring

• Fault detection and classification

• Maintenance scheduling



PO Box 210072 F 1.513.556.3390


27


Sensor:

Piezoelectric sensors from Measurement Specialties INC. are selected for hose prognosis. The sensors are placed in proper locations. Carbon Nanotube Thread (CNT) is under investigation and design for future hose prognosis.

Testing Condition:

Hose undergo pressurized cycles, until fail.

Testing Procedure:

The voltage change produced by the piezoelectric sensor is recorded by NI-9215 hardware and LabVIEW software.

Data Analysis:

Watchdog Agent toolbox in Matlab was used for data processing. Each file contains one minute data. The data analysis process is as follows:

Signal processing: FFT

Feature extraction: FFT Frequency Band; Mean; Variance; Peak-to-Peak; Kurtosis;

Peak; damping factors.

Performance evaluation: Statistical Pattern Recognition

Sensor fusion: Decision level

e) Deliverables and Results:

The confidence values (CV) calculated from Peak values, kurtosis and variance have

shown degradation trend before the hose fail. The experiment will be repeated a number

of times under the same condition to validate the result.

An alternative approach is also conducted to detect hose anomaly when hose is at rest.

The approach has shown promising results to predict hose failure. The method has

advantage in the field because it is working condition independent.

Two patent ideas are formulated and are under documentation.

2.1.9 Project Collaboration with OMRON


This project will form a value-added collaboration between two participants: the Center for Intelligent Maintenance Systems (IMS) and the Omron Corporation in Japan. The project will establish a Precision Energy Management Systems (PEMS) that will aim to reduce energy



PO Box 210072 F 1.513.556.3390


28

consumptions by linking IMS tools and techniques to Omron products, processes, and business level.


April 2008 – April 2011 (3 Years)

c) Objectives:

The vision for this project is related to three major players in the Omron infrastructure:

PRODUCT: the PEMS must be able to define, measure, analyze, and predict energy consumption per product manufactured, in such a way that each product, such as a switch or sensor, leaving the Omron process line can have an “Energy Label” that resembles the amount of energy consumed in manufacturing this product. Such an energy label can also be part of the “Eco-Label” initiative which is an integrated within Omron’s Eco-Product vision.

PROCESS: the PEMS must be able to use the developed energy metrics and parameters for detailed operational analytics. In other words, the PEMS will assess both equipment-level and process-level degradation and integrity. Such a correlation between maintenance activities on the shop floor and energy consumed through the machines and processes is an essential part of the PEMS initiative and is a large gap in today’s industry operations.

PLANT: the PEMS will enable a strategic IT-centric evolution of Omron’s QCD (Quality, Cost, Delivery) to consider energy analytics, thus inducing an Eco-IT framework for effective QCDE development and implementation.

d) Technical Approach:

Figure 14. Technical approach



PO Box 210072 F 1.513.556.3390


29

2.1.10 Project Collaboration with GENERAL ELECTRIC (GE)


General Electric is planning to develop an Intelligent Prognostic System (IPS) to monitor machine health using information available in the Computer Numerical Controller (CNC). The Center for Intelligent Maintenance Systems (IMS), as a collaborator with GE, proposes to design the offline Prognostic Analysis System (OPAS). The system will periodically conduct Fixed Cycle Feature Tests, collect data, process the data and conduct offline prognostic analysis for identifying machine degradation. Finally it should be responsible to notify users of poor machine health in a clear and intelligent way.


March 2011 – December 2011


Develop an OPAS that can learn the critical factors of machine health and analyze data collected from online Fix Cycle Feature Test, and report the analysis results.


GE's existing machine testing system periodically conducts fixed cycle tests on all six axes (X, Y, Z, A and B) of their machine tools, as well as the tool spindles. The MTConnect system collects data from machine controller signals, such as motor current and position delay, which are then saved to on offline database. This database is dedicated as a data-repository of the OPAS and buffers a “snapshot” of the machine every day for analysis. The OPAS proposed by the IMS center, will draw the FCFT data from the database for the purpose of machine tool health assessment. The data-sets drawn from the database will undergo pre-processing, including critical variable identification, segmentation, and feature extraction.

Figure 15. Sensor-less test-bed and FCFT tests



PO Box 210072 F 1.513.556.3390


30

After pre-processing, exacted data features from identified critical variables will be organized into the database for every machine. The prognostic analysis will be conducted from two aspects, anomaly detection and degradation assessment. Firstly, for anomaly detection, we will target our algorithm development towards candidate abnormal features such as spikes, shifting, periodical variation and trending. Secondly, algorithms will be developed to recognize normal conditions of individual machines and to track changes of these normal conditions. Lastly, suspected degradation will be modeled to quantitatively assess machine health. Analysis results will be reported by a Human Machine interface (in Matlab) and automatic notification email. During the analysis, information such as raw data, extracted features, detected anomalies and suspected degradation with severity index will be stored for manual inspection and historical record; otherwise the data will be discarded.


Milestone 1 –March 2, 2011: Demonstrate anomaly detection and data visualization capability

Milestone 2 – May 11, 2011: Demonstrate self-learning capability and validate classification algorithm with different machine normal conditions.

Milestone 3 – August 31, 2011: Demonstrate critical degradation modeling and assessment

Milestone 4 – October 17, 2011: Demonstrate system reporting and notification mechanism

Final project report and presentation – December 28, 2011



PO Box 210072 F 1.513.556.3390


31

2.1.11 Hong Kong’s Airport Chiller Health Monitoring


Collaborating with Hong Kong Electrical and Mechanical Services Department (EMSD) on the specific project, IMS designed and employed a systematic method called FCFT (Fixed Cycle Features Test) to identify the incipient faults of the chiller before it completely shuts down. Instead of acquiring data continuously, the chiller is run through several possible work conditions (e.g. 25%, 50%, 75% and 100% load) for two minutes every day and the overall health condition of different components will be assessed by comparing the readings acquired during these periods to the baseline.

b) Objectives:

Identifying the incipient faults of the chiller before it completely shuts down.

Developing an entire platform that integrates data acquisition, preprocessing, health assessment and visualization for implementation in the airport.

c) Summary of Technical Approach:

The data acquisition system used in this project consisted of six accelerometers installed on the housing of 6 bearings on the chiller (Figure 16). In addition, a National Instruments PXI-4472 was used to obtain data from the 6 accelerometers simultaneously. The data logging system also obtained data from the Johnson Controls OPC (Object-linking-and-embedding for Process Control) server of the chiller system through an Ethernet connection.

Figure 16. Sensor Installation



PO Box 210072 F 1.513.556.3390


32

As the working load is subject to change, the proposed method focused on identifying system behaviors during the transient period between different working loads (Figure 17). Wavelet Packet Analysis (WPA) method was applied to extract component health related features from the non-stationery vibration data obtained during these transient period. Gaussian Mixture Model (GMM) was used as the health assessment model to calculate the confidence value (CV) of the different components of the chiller. Hence, a CV (using a 0-1 scale, 0 being unacceptable or faulty and 1 being normal) was derived from both acceleration data and OPC data were converted into 0 to 1 (0-unacceptable or faulty; 1-normal) information to indicate the health condition of the chiller components. Finally, a radar chart, as shown in Figure 18, was generated to show the health condition of all the components, including the shaft, four bearings, the evaporator, the condenser, the compressor oil and the refrigerant circuit. A drop in confidence value was displayed close to the center of the radar chart, which indicated an unexpected fault was likely to happen. The method employed successfully detected the abnormal health condition of the chiller during validation.

Figure 17. Fixed Cycle Feature Test (FCFT)

Figure 18. Flowchart of Chiller Monitoring System



PO Box 210072 F 1.513.556.3390


33

d) Project Deliverables and Results:

The entire platform that integrates data acquisition, preprocessing, health assessment and visualization is compiled into one integrated user interface and is being used by the airport staff. Example interface is shown in Figure 19.

Figure 19. User Interface for Chiller Health Monitoring

2.1.12 Project Collaboration with TECHSOLVE


The project is subcontracted to the IMS Center as part of the TechSolve program called Smart Machine Platform Initiative (SMPI). IMS provides consultancy and technical assistance to TechSolve. In addition, two graduate students from the IMS Center work with TechSolve engineers to develop a health and maintenance system for a machining center test-bed.


July 2006 - June 2009 (3 Years)

c) Objectives:

Develop a scalable predictive maintenance system platform to demonstrate self-

assessment and self-prognostics capabilities for machine tool makers and users.



PO Box 210072 F 1.513.556.3390


34

Integrate the developed IMS prognostics tools and technologies with machine tool

system in partnership with TechSolve to establish a demonstration system for “Smart

Machine Predictive Maintenance Platform”

Partner with TechSolve to assist machine tools manufacturers and users to deploy the

validated tools and technologies for accelerated commercialization and leveraged

impacts.


The first major task that was done was to perform a state-of-the-art survey of machine tool

health and maintenance activities and research. The results of this task were used to identify

specific prognostics tasks that need to be addressed: tool unbalance detection, bearing health

monitoring, coolant condition monitoring, axes/ball screw health monitoring and guide

monitoring. The primary result of undertaking such a survey was to be able to identify the overall

goal of the Smart Machine Health and Maintenance System and it is summarized in the figure

below:

Figure 20.Smart Machine Platform

Tool unbalance is a critical most especially in precision machining. Although tool unbalance is constant regardless of spindle speed, the cutting forces increase with higher spindle speeds thus producing spindle vibrations. Features are extracted from vibration signals and with the use of data-driven methods; a confidence value is computed to estimate the state of tool unbalance. This confidence value is automatically sent to third-party software called Freedom eLogTM. The implementation overview is depicted in the figure below:



PO Box 210072 F 1.513.556.3390


35

Figure 21. Smart Machine Platform Implementation

Bearing health monitoring is to be demonstrated using a mock-up rig with actual machine tool bearings. Faulty bearing data are collected to train a bearing health monitoring system based on self-organizing map (SOM).


A state-of-the-art survey was generated to determine the current trend of machine health

monitoring as well as to be aware of other related health and maintenance activities.

An online monitoring system was developed to detect tool unbalance. It consists of a

data acquisition system, and a health assessment section. Using data driven methods,

the system is able to detect the state of unbalance of a tool assembly (retention knob,

tool holder, cutting tool). Furthermore, the Watchdog Agent® is able to send confidence

values (CV) to Freedom eLogTM.

A mock-up rig was prototyped as a demonstration test-bed for bearing health monitoring.

Using self-organizing map (SOM), a system was developed to detect different bearing

failure modes.

2.1.13 Intelligent Prognostic Tools for Smart Machine Platform


In order to move towards the first cut correct vision, a methodology and set of intelligent algorithms are needed to monitor the condition of the machine tool system and components. This predictive monitoring system can then assess the current and future health state of the machine and also provide guidance on whether its current state is suitable for producing parts within the desired specifications.



PO Box 210072 F 1.513.556.3390


36


June 2009 – June 2011


To develop a LabVIEW based predictive health monitoring system for machine tools. Develop suitable algorithms for predicting the future health state of spindle bearings. Develop a health monitoring system for the feed-axis. Evaluate a health model that relates tool-holder unbalance to surface quality.


Bearing RUL Modeling

The objective is to develop a set of algorithms that can estimate the health of spindle bearings and also predict the remaining useful life. The proposed approach monitors the degradation of the spindle bearings using a self-organizing map algorithm and features extracted from the vibration, motor current, and temperature signals. The vibration based feature extraction methods include the use of time domain statistics and peaks in the frequency and envelope spectrum. The health monitoring algorithm and prediction approach are highlighted in Figure 22.

Figure 22: (a) Health Assessment Algorithm, (b) Prediction Approach

(a) (b)



PO Box 210072 F 1.513.556.3390


37

0 20 40 60 80 100 120 140 160 180 2000

5

10

15

20

25

30

35

40

45

Operating Time (hours)

Health V

alu

e

Health Trend Value

162 164 166 168 170 172 174 176 178 1800

2

4

6

8

10

12

14

16

18

Operating Time (Hours)

Bearing R

em

ain

ing U

sefu

l Life

Comparing Actual and Estimated Bearing Remaining Useful Life

True RUL

Estimated RUL

Estimated RUL Lower

Estimated RUL Upper

(a) (b)

The health and prediction results using this proposed approach are shown in Figure 23; this is shown for one of the two run-to-failure data sets from the spindle bearing test-rig. The results for these two data sets have been encouraging and future work looks to further validate this method with additional run-to-failure data sets.

Figure 23: (a) Bearing Health Trend, (b) Prediction Results

Feed-Axis Health Monitoring

The feed-axis is a key subsystem on the machine tool which is subjected to a variety of operating conditions and loads, which over time can lead to the degradation of this particular subsystem. Also, a loss in performance in a particular axis of the machine tool would have a detrimental effect on position accuracy and perhaps the quality of the part manufactured. The proposed approach for monitoring the degradation of the feed-axis is provided in Figure 24.

Figure 24: Feed-axis Health Model Development Flow Chart

Acquire relevant signals: torque, position, vibration, temperature, power

Segment and extract features from the multitude of monitored signals

Evaluate various feature selection approaches to obtain the most robust

feature set for each failure mode

Train a self-organizing map for each operating condition and failure mode

Use a classification algorithm or rule base system to isolate each fault

Validate health model with seeded faults on feed-axis test rig

Refine feature selection model for improved results

Evaluate various health assessment algorithms

Evaluate classification and rules based algorithms for fault isolation

Develop final algorithm for real-time monitoring

Feed-Axis Health Model Development Feed-Axis Health Model Validation and Refinement



PO Box 210072 F 1.513.556.3390


38

0 20 40 60 80 100 1200

100

200

300

Sample #

MQ

E

Health Value of Good and Ball Nut Misalignment Fault -Using Nut Misalignment Features (No Load)

0 10 20 30 40 50 60 70 80 90 1000

5

10

15

Sample #

MQ

E

Health Value of Good and Ball Nut Misalignment Fault -Using Nut Misalignment Features (150lb Load)

0 20 40 60 80 100 1200

5

10

15

Sample #

MQ

E

Health Value of Good and Ball Nut Misalignment Fault -Using Nut Misalignment Features (300lb Load)

For developing the health model in this proposed approach, signals such as the torque, position, vibration, and temperature are further processed and features are extracted. Furthermore, a feature selection routine is used to obtain a subset of features that are best at discriminating between the good and faulty state. These selected features are used as an input into a health assessment algorithm that quantifies the level of degradation as a function of the deviation from normal behavior. Lastly, a classification algorithm can be used to isolate the particular fault that is occurring. After developing this health assessment and fault identification method, a model refinement and validation step is used to further test the model with additional data sets. Preliminary results using the proposed approach have been tested with data collected from a feed-axis test-bed. Figure 25 left figure shows the feed-axis test-rig and Figure 25 right figure shows the preliminary health assessment results. The preliminary results use data collected from a healthy system and from a feed-axis system with a misalignment fault of two levels of severity.

Figure 25. Feed-axis test-rig (Left) and Preliminary health assessment results (Right)

Tool-Holder Unbalance Health and Surface Quality Model

The overall methodology for quantifying the tool holder unbalance condition and then relating this to work piece surface quality is shown in Figure 6. In the proposed approach, vibration features related to the fundamental frequency and harmonics are used to train a health assessment algorithm. As an extension of this health model, a regression or neural network model is used to relate the surface quality to the measured inputs such as machining parameters and the tool-holder unbalance condition. Data collected from a machine tool with different levels of tool-holder unbalance is used to build and validate the health model.



PO Box 210072 F 1.513.556.3390


39

Figure 26: Tool Holder Unbalance Health and Quality Model Methodology

LabVIEW Interface

A LabVIEW based software tool for key machine tool components or subsystems, such as the spindle bearings and the feed-axis are part of the scope of this project. An example of the bearing LabVIEW monitoring program that is currently in development is shown in Figure 7; the tab shown in this screen capture is for feature plotting. There are respective tabs for feature extraction, feature selection, health assessment, and prediction as well.

Figure 27: Example of Bearing Health Monitoring and Prediction Software in LabVIEW

e) Project Deliverables:

LabVIEW based monitoring system configured for monitoring the critical machine tool components and subsystems.

Documented report that highlights the methodology and results for each monitored machine tool component or subsystem.



PO Box 210072 F 1.513.556.3390


40

2.1.14 Project Collaboration with SIEMENS


This work was proposed to Siemens TTB in early 2008 and is currently in the final stage of IP agreement negotiation. The work is initially defined to have one student from the IMS Center working in Siemens TTB for 6 months (subject to change).


2008


The goal is to develop a reconfigurable “Plug-n-Prognose” black box which can be easily and effectively used for practitioners to assess and predict the performance of the critical components (bearings, gearboxes, and motors) of the rotary machinery (see figure below). Visualization tools are provided to present the information of the components’ health which can be further utilized for precision maintenance decision making.

Figure 28. reconfigurable “Plug-n-Prognose” black box


The innovative idea of the work will be that the product can automatically “sense” different input signal characteristics which will be linked to the input of the selection tool. The selection tool will then automatically prioritize the prognostics algorithms which will be used thereafter. After some pre-configuration, the product can provide accurate prognostics information of rotary machinery components with minimum human intervention.


The deliverables will be a demonstrable prototype of the autonomic prognostics platform for rotary machinery. The platform integrates both hardware (PC/104 industrial PC, DAQ board and



PO Box 210072 F 1.513.556.3390


41

accessories) and software (IMS Watchdog Agent® Toolbox) by developing autonomic tool selection method, reconfigurable prognostics architecture and a user-friendly interface to enable remote and embedded health assessment and prediction.

2.1.15 Project Collaboration with Hiwin, Taiwan

a) General background:

The increasing feeding rate and high acceleration value have improved the machine tools efficiency and accuracy and also reduce the machining times. The each component of the machine tools can directly affect the overall precision of the machine tools system. The critical components can be defined based on their criticality level, low failure frequency but high impact. The condition-based monitoring can enable the manufacturing factories to observe the degradation of each critical component and calculate their the health value, therefore the overall health condition of the machine will be evaluated based on the health condition of each component. This health indicator like health value, confidence value can provide and predict the information of the machine’s health condition. So the predictive analytical algorithms provide significant improvements of adding PHM to traditional maintenance schemes. Machine data is effectively transformed by PHM algorithms into valuable information that can be used by factory managers to optimize production planning, save maintenance cost and minimize equipment downtime.

The ball screw is the one of the most important elements for industrial machinery and precision machining. Its primary feature is to convert rotary motion to linear motion with high accuracy, reversibility and efficiency. However the ball screw performance degrades over time after repeated usage, which can impact the feed-axis system and undesirable loss of position accuracy. Thus, the accuracy of ball screw is the prior concern based on the quality requirement of high precision performance machines. By acquiring information of the current health state of the ball screw and its degradation trend, it will save costs in maintenance and avoid products with low quality and downtime due to unexpected ball screw failures if proactive maintenance can be scheduled before the machines loss accuracy, or even failure happens.

b) Project duration:

Phase I: January, 2011 ~ July, 2012

Phase II: January, 2013 ~ March, 2014

c) Project objectives:

This research focused on further establishing a strategy and technical architecture for prognostics and health management (PHM) of ball screw. By analyzing the vibration data and temperature data acquired from the different location on the ball screw, this study’s target is to develop and apply a PHM strategy and technical approach for early detection of failure and Remaining-useful life prediction. Frequency and time domain features are extracted to evaluate the degradation. Neural Network (NN) is applied to map the information gained from external



PO Box 210072 F 1.513.556.3390


42

sensors with position loss. So far, the position model is established using data collected from one test and validated from another one. The method showed reasonable results, but it also needs more testing data to refine the model and increase the prediction accuracy.


In order to understand the ball screw’s degradation process, experiments have been conducted in both the Hiwin factory and also the IMS center lab. Specific highly stress tests have been designed to accelerate the ball screw’s degradation. Learning from both sensor-less and sensor-rich fault diagnosis tests in 2011, it is quite obvious that the external sensors can collect more information and directly target the failure modes and location when compared with only motor controller data. Thus, extra sensors like accelerometers and thermocouples are installed on certain key locations to monitor its performance degradation. At the same time, the linear encoder is used to measure the position loss in order to validate and benchmark the health monitoring methods.

The overall methodology is developed under Matlab® environment. For the ball screw case, it has its own specific movement pattern, so a systematic method is tailored to analyze the ball screw data, as seen in Figure 29.

Figure 29. Ball Screw Health Monitoring System Flowchart


In the smart factory, all the physical entities in the factory are connected together physically and virtually. In the physical space, the sensors are utilized to measure and detect the occurrence of

BallScrewHealthAssessmentandFaultDiagnosis

SetupDataAcquisi onandSignalSet

DataPreprocessing

DataQualityCheck

SelectToolsforhealthassessmentandfaultdiagnosis

!

N

Y

PreliminaryAnalysis

OutlierDetec on

N Y

DiagnoseDegrada onPa ern

FailureModeIden fica onandEvalua on

LifeDegrada onPredic on

RULPredic on(Posi onErrorEs ma on)

stopN

Y



PO Box 210072 F 1.513.556.3390


43

a fault or anomaly, and isolate the location of the fault and identify the failure types. Beside the sensor signals, the outputs from the machine controller like current, voltage, power consumption can also be used to reflect the machines’ overall health degradation. These physical entities are connected in the cyber space through Internet. The sensor signals are collected through data acquisition hardware, and either stored locally or uploaded to cloud server. After storing the data, pre-processing techniques are employed to check the data quality and increase the signals’ signal-to-noise ratio (SNR), such as the data filtering, quantization, etc. Data segmentation is used to select the interesting segment from the input signals. Furthermore, the features (health indicators) are extracted, and health value is calculated over time to evaluate the behavior of the machine components and detect incipient signs of failure even before the actual failure event. Beside of this data-driven method, the physical model and experiment knowledge are also applied to fully understand the performance of these components, like the FEA model, mode analysis, etc. In the cyber space, all of these models are integrated to convert the data to information, with which the customers can schedule the timely maintenance before the failure.

In order to model the components’ behavior quickly and effectively, the accelerated degradation test in the lab is utilized in the machine reliability analysis, the different stress factors are screened and the important stress factors are picked according to the quantization matrix. After that, the characteristics tests are conducted to identify the initial condition. Thus the data collected from multiple degradation process under different stress levels can generate the degradation pattern for the component, thus the time conversion between these different degradation pattern will be validated, which can assist to identify the degradation trend in the actual industrial application. After establishing the degradation model, this PHM model will be integrated in the cyber space. The data collected from the machine components will be processed through this PHM model and thus a virtual component is modeled in the cyber space, so it can simulate and predict the health pattern of the machine component in the reality.



PO Box 210072 F 1.513.556.3390


44

Figure 30.Smart Machine Component Prognostics Scheme

2.1.16 Project with INSTITUTE FOR INFORMATION INDUSTRY (III), TAIWAN

a) General background:

With the growing maturity of the Information and Communications Technology (ICT), machining and manufacturing sectors have been accelerating the pace to integrate with the related ICT techniques, so as to enhance the competence among global competitors. Since the number of machines being monitored increases, the volume of the data being collected at every level of machines/manufacturing lines/factories to be transferred and processed becomes so enormous that the conventional techniques need to be improved. Also, due to differences between machining processes for different products, the parameters and physical characteristics that need to be monitored vary from case to case. Therefore, in order to efficiently collecting diagnostic information, a dynamic feature extraction system is to be designed and integrated with the current machine tool monitoring platform. This upgrade of the software will enable the data collection system to dynamically select the useful parameters related to machine health, and enlarge expert system library to alleviate the burden of communications and increase the efficiency of value-added ICT service.

PhysicalSpace

CyberSpace

CyberSpace:• BallScrewTime

Machine

• BallScrewClusters

• HealthEvalua onandLifePredic on



PO Box 210072 F 1.513.556.3390


45

Figure 31. Framework of dynamic condition-based feature extraction

b) Project duration:

April, 2013 ~ November, 2013

c) Project objectives:

The developed real-time dynamic feature extraction approach is to be integrated with the intelligent machine tool DAQ platform “ServBox”, which is a product of Institute for Information Industry. The objectives of this research include:

Build a knowledge library of mechanical component health diagnosis that consists of (1) a summary features to be extracted, (2) relations between features and effectiveness of features for diagnosis and, and (3) quantified evaluation of the effectiveness;

Design a hierarchical structure for dynamic feature extraction.

d) Summary of technical approach:

The system includes 3 main function units:

1) A knowledge base for critical machine components, their failure modes, and related data resources/features. It is constructed through expertise interviewing and literature survey and based on quality function deployment (QFD).

2) A decision tree for feature priority will be constructed automatically based on Bayesian probability learning from machine condition and prior knowledge. The execution flow defines

Dataacquisi ondevice



PO Box 210072 F 1.513.556.3390


46

most critical features at the first level (so called Pivot Features) and secondary features in lower levels.

3) In routine monitoring tasks, under normal conditions, DAQ only calculates Pivot Features and analyzes them to make initial judgments of whether anomaly happens. If any abnormal condition is detected, DAQ will further go through lower levels, calculate and evaluate secondary features, and eventually infer root cause of the problem.

By incorporating the proposed mechanism, DAQ will be able to operate with minimal computational resource under normal conditions. The overall procedure is presented in the figure below. Besides, the PHM system can react and perform diagnosis promptly when anomaly happens. With such improvement, the future DAQ and PHM system together will be able to perform real-time health diagnosis and prognosis in a more bandwidth and computational efficient way.



PO Box 210072 F 1.513.556.3390


47

Figure 32. Overall approach flowchart for dynamic feature extraction

e) Project deliverables and results

Project deliverables:

Library of mechanical component health diagnosis, including feature extraction algorithms, and quality function deployment (QFD) table;

Component(Sensor(Feature/Table/

Sensors/ Controller/ Vibra5on/ Power/ …/Current/ Posi5on/

Features/ Mean/ Accuracy/Variance/RMS/ Kurtosis/ Sha@/freq.//Bearing/freq./Mean/Max/ …/

Componen

ts/&/

Problems/

Spindle/ L/ NA/ NA/ M/ L/ H/ NA/ L/ L/

…/

Bearing/ L/ NA/ NA/ M/ M/ L/ H/ L/ L/

Machine/Tool/ M/ NA/ NA/ M/ M/ M/ NA/ H/ H/

Tool/Holder/ NA/ NA/ NA/ M/ M/ M/ NA/ NA/ NA/

Feed/Axis/ L/ M/ M/ M/ L/ M/ NA/ L/ L/

Servo/Motor/ H/ L/ NA/ L/ L/ L/ NA/ H/ H/

Comp./Power/ Low/ Medium/ Medium/ Medium/Medium/ High/ High/ Low/ Low/…/Comp./Time/ Medium/Medium/ Medium/ Medium/Medium/ High/ High/ Low/ Low/

Storage/ Low/ Medium/ Low/ Low/ Low/ Low/ Medium/ Low/ Low/

H (high, or strong), M (medium) or L (low, or weak) and NA (not applicable)

Score/values:/NA/=/0,/L/=/1,/M/=/3,/H/=/9/

Identify critical

problems and related

features

Construct

Component-Sensor-

Feature table

• Spindle

• Bearing

• Machine

Tool

• Tool

Holder

• Feed

Axis

• Servo

Motor

• Current

average

• Vibration

RMS

• Power

average

• Frequency

Amp.

• …

Problems / Components Related Features

First layer construction

Second layer

construction

Human machine

Interface integration

!

» By incorporating the proposed mechanism, DAQ will be able to operate with minimal computational resource under normal conditions.

» The PHM system can react and perform diagnosis promptly when anomaly happens.



PO Box 210072 F 1.513.556.3390


48

Feature extraction hierarchical levels and the algorithm to build them;

A published research article.

Results:

The gearbox data from the PHM 2009 Data Challenge were used for method validation. The model has achieved 100% accuracy for health assessment, 100% accuracy for gear and bearing fault diagnosis, and 100% detection rate (with 12% false alarms) for shaft fault diagnosis. In terms of data processing speed, the algorithms consumed on average 0.24 seconds to process 1.5 length of data under healthy conditions, and 0.92 seconds under faulty conditions when diagnostic feature extraction was triggered. The testing results well demonstrated the capability of the proposed method in optimization of computational recourses while maintaining fault classification accuracy.

Table 1. Sets of experiment with different health conditions used for model training and testing

Figure 33 Health assessment and fault diagnosis result using dynamic feature extraction

0 100 200 300 4000

0.2

0.4

0.6

0.8

1

Hea

lth

Va

lue (

CV

)

Time(s)0 100 200 300 400

0

0.2

0.4

0.6

0.8

1

Pro

ba

bili

ty o

f G

ea

r F

au

lt

Time(s)

0 100 200 300 4000

0.2

0.4

0.6

0.8

1

Pro

bab

ility

of B

ea

ring

Fa

ult

Time(s)0 100 200 300 400

0

0.2

0.4

0.6

0.8

1

Pro

ba

bili

ty o

f S

ha

ft F

au

lt

Time(s)

Baseline 1

Chipped Gear&Eccentric Gear

Ecentric Gear

Baseline 2

Broken Gear&Ball defect Bearing

Inner Race Defect&Bad Keyway

Baseline 3

Outter Race Defect&Imbalance Shaft



PO Box 210072 F 1.513.556.3390


49

2.1.17 Project Collaboration with FMTC


This project was a collaborative work between IMS center and FMTC on developing novel winding fault detection techniques in 3-phase induction motors. Among the different faults that may develop in an induction motor, winding faults are the second most occurring faults after bearing faults. Winding faults can cause severe damage to the induction motors if they progress. However, if detected early, they can be fixed with minimal cost and a major downtime and repair cost can be avoided. Over the course of this project, IMS center worked with FMTC and successfully developed early winding fault detection techniques based on current, voltage and vibration signals.


June 2013 – June 2014


Objective 1: Development of an early winding fault detection technique based on current and voltage signals

Objective 2: Development of an early winding fault detection technique based on vibration signals


FMTC designed and built a test-bed consisting of an induction motor, a variable frequency drive and a magnetic brake for applying load on the motor. Winding faults of different types and different levels of severity were induced in the motor using shorted circuits and a variable resistor. Current and voltage signals at three phases were collected along with vibration and motor speed signals.

For analyzing the current and voltage signals, a signal processing technique was developed to preprocess the signals and compensate the effect of the variable frequency drive on the signals. Afterwards, three different techniques including negative-sequence impedance, voltage mismatch approach and Park’s vector approach were tested on the data. The results from the three techniques were consistent and successfully detected the incipient winding anomalies in the induction motor.

For analyzing the vibration signals, a signal enhancement technique based on kurtogram was developed. After enhancing the signals using Time Synchronous Averaging (TSA) and retaining the frequency band of interest, envelope analysis was applied to detect the winding fault related impacts in the vibration signature. The results were consistent with the results obtained from motor current and voltage signals and proved to be capable of detecting winding faults in early stages.

The results from this project were presented in two international conferences.



PO Box 210072 F 1.513.556.3390


50

2.1.18 Project Collaboration with Korea Aerospace University (KAU)


This project was a collaborative work between IMS center and KAU for developing a Prognostics and Health Assessment (PHM) framework and algorithm for material handling systems. Material handling systems consist of a multiplicity of components used for storage, retrieval and handling of goods and equipment. The objective of this project was to monitor their critical components to reduce the costs of downtime and maintenance and improve the productivity of such systems.


Phase 1 (Survey): July. 2014 – Dec. 2014 (6 months)

Phase 2 (PHM System Development): March 2015 – Aug. 2015 (6 months)


Objective 1: A survey paper and a report on the PHM methodologies for critical components of material handling systems

Objective 2: Development of PHM systems for critical components of material handling systems

Objective 3: Development of a Cyber-Physical framework for material handling systems


During the first phase of the project, a comprehensive survey was performed on the health assessment and monitoring of the critical components and sub-systems of material handling systems including conveyor systems and shuttle battery. The paper was published in the International Journal of Advanced Logistics.

For the second phase, IMS center and KAU are collaborating to develop methodologies and algorithms for PHM of material handling systems. A test-bed was setup for testing the batteries used for shuttle and experiments were performed on the batteries during various charge/discharge regimes to develop a model for the degradation of the batteries. Currently, the tests are still ongoing and IMS is working to develop a degradation model for batteries to estimate their remaining useful life. KAU is designing a test-bed for acquiring data from other critical components of typical material handling systems. IMS center will be analyzing the data to build degradation models for such components and sub-systems. Moreover, IMS center will develop a framework for building a Cyber-Physical System (CPS) for material handling systems.

2.1.19 Project Collaboration with Cosen


Machine Tool Condition Monitoring (TCM) is an effective way for achieving optimal utilization of machine tools, with which the best time for replacing a worn-out tool can be determined by



PO Box 210072 F 1.513.556.3390


51

maximizing the usable time of the tool while maintaining the desired product quality. Moreover, TCM can help schedule predictive maintenance, reduce unexpected downtime and potential safety issues.

In factories, manufacturing processes start with sawing large pieces of material into designated sizes. Due to the up-stream nature of the sawing process, the quality and speed of sawing affect the entire production and any error could be propagated to the following steps and result in bad quality products. On the contrast, performing accurate cuts require slower cutting process, which hampers the productivity of the whole manufacturing line. Therefore, while it is not possible to maximize these two requirements at the same time, an optimal balance between surface quality and speed has to be achieved. In such situation, the health status of band sawing machine and its components play significant role in both speed and quality of cuts. Thus, availability of exact health information from the machine will provide significant help on maintaining the cutting speed and accuracy.


July 2013 to July 2015


To develop health assessment and prognosis algorithm for key components of sawing machine including blade, motor, hydraulic systems and coolant systems

To deploy the developed solution on a user-oriented cloud platform with user interface and online data acquisition system

d) Summary of Technical Approach

Figure 34. Cloud setup for distributed sawing machines



PO Box 210072 F 1.513.556.3390


52

Three band saw machines have been setup at different geographical locations. The overall system setup is shown in Figure 34. Data has been acquired from machines through both add-on sensors and controller signals. In addition to the add-on vibration, acoustic emission, temperature and current sensors, 20 control variables such as blade speed, cutting time and blade height have been pulled out of the PLC controller to provide a clear understanding of each machine working status.

Table 2. Data table structure of cloud database for machine prognostics

Database Sections Column Names

Primary Keys Machine_ID Date_Time

Cut_Cycle_#

Working Regime Blade_Speed Material_Type

Cutting_Rate …

Health Feature Vib_RMS Vib_Kurtosis

Frequency_Band_Energy_1 Frequency_Band_Energy_2

AE_RMS …

Health Condition Confidence_Value Health_Stage

At each test location a local industrial computer, as the cloud agent, performs feature extraction and data preparation. The feature extraction consists of extracting conventional time-domain and frequency domain features such as RMS, kurtosis, frequency band energy percentage, etc. from vibration and acoustic signals. Calculated features along with other machine state data is being sent through Ethernet or Wi-Fi network to the cloud server where the data is managed and stored in the database. The prognostics table structure in the cloud database is showed in detail in Table 2.

On the cloud, a recurring program as data watcher agent keeps track of incoming data and upon arrival it executes adaptive prognostics in real-time. At this level, based on machine working state data, current working regime of the machine has to be determined by an adaptive method. The method determines current working regime by comparing it with historical machine state clusters and if required will create a new working cluster for current state. After working regime determination, machine health values and health stages are adaptively generated based on the overall trend and noise level of features from each machine (Figure 35). The health



PO Box 210072 F 1.513.556.3390


53

stages are then considered as states by the algorithm for state based prediction using a Markov Chain Monte Carlo method. Meanwhile, users are also alerted when a new stage is created as a sign of further blade condition deterioration.

Figure 35. Adaptive health state recognition for blade condition


A cloud based condition monitoring and prognostics platform has been delivered to Cosen Saws, with the algorithms deployed in cloud server, WEB based user interface, SQL database and Advantech based real-time DAQ system. The platform is being evaluated by Cosen customers.

2.1.20 Project Collaboration with Daniel Drake Center for Post-Acute Care and

Beechwood Home

a) General Background

Using a range of integrated sensor systems for acquisition of various health and activity related data, the IMS Center is developing information-driven patient profiles that can aid medical professionals in clinical decision making and resource scheduling for patients. The IMS Center is working with long-term care facilities like the Beechwood Home and rehabilitative centers like the Daniel Drake Medical Center for Post-Acute Care to gather necessary data to build these systems.

b) Project Duration

August 2013 – TBD



PO Box 210072 F 1.513.556.3390


54

c) Objectives

Create smart technologies such as the smart awareness wheelchair and wheelchair tracking system and utilize wearable gait sensors to instrument stroke patients.

Enable these data sources to log information to be extracted as features for patient health data.

Develop an adaptable peer-to-peer model to be able to predict patient outcomes based on historic data of patients with similar attributes.

Integrate the health assessment and prediction from these patient models to decision support visualization system to communicate information to health professionals.


In industrial applications, various signals are taken to monitor, assess and predict conditions of components. These same types of sensors can be fitted to a person or healthcare equipment to monitor conditions of the human body and provide insight into a person’s conditions in the same way the sensors do with machine prognostics.

Signal processing and feature extraction methods vary greatly from sensors. Some require basic time-domain or frequency-domain techniques, whereas multidimensional GPS data requires further dissemination to define location areas and event sequences.

For some data, a simple anomaly detection will be employed based on historic data and/or expert-defined thresholds. For example, an alert will be provided to the healthcare staff if a patient has fallen from their wheelchair or if they haven’t moved from an area in an unusual amount of time.

For health data, patients will require classifiers based on expert-knowledge for patient conditions to train the variable model (i.e. ranking from no activity to normal activity). If a patient begins a rehabilitative regiment with little to no muscle activity, testing from their initial visit may be able to predict when they will reach normal ambulatory movements and what treatments may give them the best outlooks. Similarly, if a noticeable trend appears that is indicative of declining health, medical professionals can intervene with proper care of prepare proper resources to assist a patient or resident.



PO Box 210072 F 1.513.556.3390


55

Figure 36. Applying Watchdog Agent ® Tools for Health Care Monitoring

e) Project Deliverables and Results

Anomaly detection system for long-term care facilities to alert staff members of residents who may require assistance, while still granting them the freedom and independence that the facility promotes.

Improvement of peer-to-peer model to be adaptable to patient variable attributes.

Integration of health model and patient outcome analyses to visualization system for healthcare providers.



PO Box 210072 F 1.513.556.3390


56

2.1.21 Project Collaboration with Woodward Inc. (Motor)


Due to its inherent efficiency and reliability, brushless DC (BLDC) driven actuation systems are widely used in a variety of industries such as aerospace, electric transportation and industrial positioning. However, it is inevitable that various types of faults can develop in the actuator either from the BLDC motor or geared positioning systems. Numerous research efforts have been made to detect and diagnose BLDC motor faults as to prevent consequent downtime caused by actuation failure. It is critical not only to identify the faults when they occur, but also to predict failure of the actuation system against some predefined failure thresholds. In particular the project collaboration between IMS and Woodward strives to predict gear failure in BLDC electric actuators using existing feedback sensors rather than adding cost with the inclusion of accelerometers.


June 2012 – Present


Develop an understanding of the Simulink model representing the controls and physics of the actuator. The sophisticated model provided the capability to simulate the response of the system with certain parameter changes such as friction or backlash, which might represent a certain level of degradation of the gears in real application.

Considering the cost and vulnerability of accelerometers, it was desired to explore the feasibility of using only the existing feedback sensor technology of the actuator system for diagnosis and performance prediction.

Develop a method that is formulated within a "sensor-less" environment and makes full use of existing on-board sensing information, which provides the possibility of a closed-loop control system for life management.


Through the implementation of a test bed within Woodward Inc., the proposed method is able to use run-to-failure endurance data to extract useful information from only current and position signals about the gear wear within the actuators. Over the duration of the tests, various time and frequency domain features were extracted and examined for proper representation of the degradation developing within the actuators. Following this, it was desired to use the features to generate a performance indicator, this was done by training a Self-Organizing Map (SOM) with healthy data features and using the Minimum Quantization Error (MQE) from test data as a performance indicator. This provided a clear visualization of the health degradation as the tests continued and were completed. This also offered a means to create a model for prediction.



PO Box 210072 F 1.513.556.3390


57

Using a general regression neural network, the project successfully demonstrated the ability to predict gear failure well before actual failure.



PO Box 210072 F 1.513.556.3390


58

2.2 Fundamental Research Projects

2.2.1 A Systematic Methodology for Data Validation and Verification for

Prognostics Applications


This is a NSF fundamental research project that the IMS center at University of Cincinnati, University of Michigan and Missouri University of Science and Technology collaborate to explore new ways to control and therefore assure the quality of data that is used for prognostics applications. Currently, significant effort is spent on data processing and developing and deploying prognostics algorithms, often only to discover, in the end, that the initial data collected was flawed. The research question here is how to assess a dataset before investing significant resources on analysis.

This research provides answer to the following questions:

Is the acquired historical data good enough for PHM decision making purposes?

Does the test-bed have efficient number/types of sensors installed?

How much data for testing needs to be acquired to solve the problem?


July 2010 – July 2012


To design a PHM focused methodology for design of data collection systems

To establish a multi-dimensional index, or a number of tests, to represent the quality of prognostics datasets

To provide a systematic methodology to extract data behavior such as linearity, randomness, dynamism and dimensionality

To provide a systematic methodology to extract data structure based on working regime, physics of the system etc.

To recommend appropriate algorithms based on the data behavior and structure

To research data reconstruction methodologies to improve the quality of PHM datasets


Targeting at two types of PHM analysis: fault health assessment and diagnosis, this research work focuses on evaluating and improving the quality of training data by identifying outlier instance and eliminating inefficient feature-set prior to diagnostic classification modeling. Thereby qualified training dataset will result in a classifier with higher predictive accuracy.

The proposed method addresses the following data quality issues for diagnostic classification modeling:

Unknown or mistaken identities of the data

Inappropriate selection of sensors or feature sets

The developed method consists of the following tasks:



PO Box 210072 F 1.513.556.3390


59

Establish quantitative metrics to validate that the training data are suitable to be used to in classification modeling for multiple failures

Filter out the outliers instances and inefficient features/sensors

The Graph spectral analysis based approach is shown in Figure 38.


The deliverables of this project include the data verification and validations methods that quantitatively assess the process of data collection, the quality of the recorded sensor data, repairs damaged or corrupted data points, and aid selection of a suitable prognostics tool/agent.

f) Future Work:

The next step in this research is to propose a method for determining the quality of the data for prognostic purposes by defining prognostic metrics for different types of data and different applications.

Figure 37. Overview of the clustering tendency method

(a) Traditional modeling procedures

(b)Automatic data quality assured modeling procedures

Data

Feature

extraction

Feature

selectionFeature

vectors

classification

algorithms

Vibration moments,

Signature

frequency,…

Multiple failure

conditions

data model

Training data



PO Box 210072 F 1.513.556.3390


60

Figure 38. The clustering tendency approach based on spectral analysis

2.2.2 Hybrid Modeling of Equipment Efficiency and Health for Advanced

Servo-drive Presses


This project focuses on the modeling of new generation servo-press machines. The Center for Intelligent Maintenance Systems (IMS) at University of Cincinnati and the Center for Precision Forming (CPF) at the Ohio State University will collaborate to research and integrate the process model of new generation of servo-press machines (CPF expertise) with the data-driven prognostics and predictive maintenance models (IMS expertise). Figure 39 illustrates the methodology for modeling of servo-press machines.

This project is fully granted by National Science Foundation and aims to improve the technology of servo-press machines. For modeling the machine, this project needs to have access to a servo-press database. Several types of sensory data will be acquired from the test-bed e.g. strain, vibration and current.

The newly developed servo-press machines are produced by Aida and Komatsu and are being used by Honda and Toyota. In this step, IMS and CPF are communicating with the companies to get access to a test-bed and/or a database.


July 2010 – June 2012

Ne

w R

ep

rese

nta

tion

0 20 40 60 80 100 120 140 160 180 20020

40

60

80

100

120

140

160

180

1:rejects H0 at

the 2.5 % significance level

2: Accept H0

3: Accept

randomness(H0)

4: the small

number of instances

A Quality Evaluation of Data

Partitions

0 10 20 30 40 50 60 70 80 90 1000

10

20

30

40

50

60

70

80

90

100

sample index

sam

ple

index

-1

-0.5

0

0.5

1

-1

-0.5

0

0.5

1

0

0.5

1

Spectral Decomposition Data Partitioning

1

2 3

4

Dis

sim

ilarit

y S

pe

ctru

m

Feature Selection

Training Data

Quality

Assured

Training Data

Outlier Detection

0 10 20 30 40 50 60 70 80 900.5

1

1.5

2

2.5

3

3.5

4

4.5

instance indices

Lo

ca

l O

utlie

r F

ac

to

r (

k/4

ne

are

st n

eig

hb

ou

r)

outliers

5 10 15 20 25 30 35

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

Feature Indice

H1

Feature Scores



PO Box 210072 F 1.513.556.3390


61


One of the main objectives of this project is to develop a physical model of the servo-press machines to achieve lower energy consumption, higher stroke speed and higher product quality. This part will be done by CPF. The second task will be done by IMS and will help to investigate Condition Assessment and Diagnosis-and-Prognosis Modeling. This model will help to improve the reliability of the machine and lessen the costs associated with the machine down-time and maintenance. The last part of the project is devoted to the integration of the two mentioned models which will help to validate and instrument the developed models for application in servo-press machines. The overview of the project is shown below.


The summary of the proposed technical approach is as follows:

Segmentation of the stamping cycle based on the strain and its derivative; the process in divided into 3 stages: pre-contact, press and after-press, as shown in Figure 40 (left figure).

Feature extraction; for first two stages, 6 time domain features are extracted, and for the vibration stage, 1 feature is extracted using continuous wavelet analysis.

Fault detection & classification

SOM for fault classification. The obtained results are shown in Figure 40 (right figure).



PO Box 210072 F 1.513.556.3390


62

Figure 39. The general approach for hybrid modeling of servo-drive presses



PO Box 210072 F 1.513.556.3390


63

1350 1400 1450 1500 1550 1600

0

1

2

3

4

5

Data records (index)

Am

plit

ude

Strain signal of stage 1 and 2 from 4 different cases

Misfeed stage 1

Misfeed stage 2

Normal stage 1

Normal stage 2

TooThick stage 1

TooThick stage 2

Slug stage 1

Slug stage 2

U-matrix

0.196

1.45

2.7

SOM 24-Jun-2011

Labels

3

3

3

4

1

4

1

4

2

1

1

4

Figure 40. (Left) Some of the vibration signals along with the applied segmentation (different colors) are shown. (Right) the obtained results using SOM


This project helps the industry to improve the quality of the products by improving the performance of the machine. The cycle time of the machine and also its energy consumption can be reduced by improving the design and control system of the machine. Also, the developed model can be implemented in the machines which are currently in use in order to improve machine reliability and reduce the costs associated with the down-time of the machine and the unpredicted failures.

2.2.3 Battery Prognostics


Current advances in energy-storage technology have led to the development of new batteries with higher capacity, smaller size and weight, and shorter charging cycles. The emergence of the new battery technology has enabled the realization of new applications that would have been a dream until recently, in-particular new electric and hybrid-electric vehicles. With these new advances in battery technology, the battery has transformed from a primitive energy-storage device to a more advanced energy-storage device.



PO Box 210072 F 1.513.556.3390


64


June 2010 – June 2011


Currently at IMS there are different ongoing activities to study and develop prognostic and diagnostic solutions for different types of batteries. The activities include researching and developing battery prognostic techniques; developing a battery prognostic simulator to speed up the development and deployment of battery prognostics; and setting up a flexible battery prognostics test bed to test new battery technologies. The overall prognostics methodology is shown in the Figure 41 below.

Figure 41. Battery Prognostics Methodology


The first step would be data acquisition. The second step would be extracting features from different regimes. This step is a very important step in designing battery prognostic solutions. Usually the raw data available from the battery will contain specific behavioral patterns that can be seen in features. From the raw data many several features could be extracted, but not all the features will be degradation related features. Degradation related features, are specific features in the raw data that will change only for degradation reasons. Currently several features are being extracted from three different regimes (charge, discharge and testing cycle). Some of the features extracted are as follows:

1. Rate of voltage buildup during charging cycle.

2. Rate of current drain during charging cycle.

3. Rate of battery temperature change during voltage buildup

4. Rate of battery temperature change during current drain

5. Maximum battery temperature per charge cycle

6. Maximum battery temperature per discharge cycle

7. Internal battery resistance



PO Box 210072 F 1.513.556.3390


65

After feature extraction, different health assessment tools could be utilized to assess the health of the battery in each regime. The selection of the health assessment tool will depend on the available training data and application type. It is necessary to note that a degrading battery might have different degradation patterns in different regimes, so it is very important to assess the health in different regimes. After generating the health value in each regime, a risk chart can be utilized to determine the overall health of the battery by fusing the fusing the health values in each regime.


A test bench in IMS center could generate the kind of data specifically needed for prognostics study of the batteries. With this system it would be possible to test different battery technologies while implementing real life abusive charge/discharge cycles, study the battery behavior and train the model based on the test data in service life of the battery. Figure 42 shows two battery test beds in IMS.

Figure 42.Battery test beds in IMS

Being flexible, it could also simulate fault-induced charge/discharge cycles for model training purposes. A few examples of these cycles are high-C charging, under –voltage/over-current discharging, partial charging…. Another application of this test bench is validation of the developed battery prognostic models and fine tuning of them over time. This will lead to a better understanding of the model in the future.

It will improve the precision of current battery RUL estimation and prediction by integrating IMS tools for automatic extraction, timely delivery, and appropriate visualization of relevant equipment information for predictive maintenance, battery application optimization, service consulting, etc. This will address the current technology gaps that are not being met by the existing competition, as well as position the company’s products to compete in the global market.



PO Box 210072 F 1.513.556.3390


66

2.2.4 Remaining Useful Life Prediction of In-vehicle Battery Pack


The project is a value-added collaboration between the Center for Intelligent Maintenance (IMS) and Shanghai Jiao Tong University, China. The research will find a systematical way to precisely and smartly predict the remaining useful life of the in-vehicle battery pack, manage the mobility of the electric vehicles, and develop soft computing algorithms based on the cloud computing technology. The integrated smart battery system will drastically improve the drive experience for the electric vehicles.


July 2010 – July 2013


To find the degradation curve of the battery pack for the purpose of RUL prediction To find the methods to extract the features of the road conditions and driver’s driving styles To develop an integrated smart battery prognostic system To develop a cloud-computing-based smart mobility maintenance system for the electrical

vehicles


As shown in Figure 43, the whole smart mobility system involves the connections between three entities: electric vehicle, road/locations, and user/driver. In the middle of each pair of entities, there can be an intelligent analysis to either improve the user experience or save the energy. For example, the battery management system indicates the status of the battery pack and how far the electric vehicle can go. The cloud-based telematics system locates the vehicle in real-time and given the charging stations around, it can dynamically updates the mobility information. The driving behavior analysis system can statically record the routes user drives and the target places as well, and then extract the pattern of the routes and driver’s driving pattern.

As shown in Figure 44, the approaches can be categorized into three major key independent methods. The first category of methods is about the battery prognostics. Features are extracted from the real-time signals of voltage, current, and temperature. Then a modified particle filter algorithm is used to predict the future performance of the batteries. The second category is to extract the energy consumption features of the road conditions such as vibration, temperature, etc. when the electric car running on the road. The third category is to the self-learning algorithm to describe the user’s driving pattern.



PO Box 210072 F 1.513.556.3390


67

Figure 43.Overview of the smart mobility technique scheme

Figure 44.Technique approaches

e) Project Deliverable:

Degradation pattern analysis of the battery pack for the purpose of RUL prediction The methods to extract the features of the road conditions and driver’s driving styles An integrated smart battery prognostic system A cloud-computing-based smart mobility maintenance system for the electrical vehicles



PO Box 210072 F 1.513.556.3390


68

2.2.5 Virtual Bearing


As a main cause of failure in machinery, rolling element bearing fault has been widely researched and numerous condition monitoring and diagnostics methods have been developed. However, due to the highly dynamic nature, multiple failure modes and extensive usage of rolling bearings, it is still urgently needed to develop an effective prognostics system that can be reliably applied in real-life situations. Furthermore, most bearing PHM related researches are based on lab-test data, which is acquired from bearings with artificially introduced faults or accelerated life test. Run-to-failure bearing data collected in-field is needed to validate current PHM models and develop more accurate and applicable models in future.


Dec 2010 – Dec 2012

c) Project Objectives and Technical Approach:

In order to solve the problems of real-life prediction and insufficient data sources, a virtual bearing system based on cloud-computing will be developed as shown in Figure 45. The core of the system is a knowledge base containing baseline, failure modes and degradation pattern information for bearings regarding different categories such as usage and operating conditions. When new sets of real-life bearing data are uploaded by industrial users, best matching units from the knowledge base will be selected through pattern matching and data mining approaches and the most suitable model will be used to perform data analysis. Then services such as the health condition and/or prognostics results will be returned to users. Furthermore, the uploaded bearing data will be used by the system knowledge base to adjust and update its corresponding information and models. Eventually, through this type of self-learning and self-expansion mechanism the maximum potential of PHM accuracy and reliability for rolling element bearing can be achieved.

Figure 45. system schematic of virtual bearing project



PO Box 210072 F 1.513.556.3390


69


Accurate bearing RUL prediction considering different degradation patterns Bearing models for prediction according to different usage (load/speed) and working

condition (constant/transient) A virtual bearing system based on cloud-computing for health management of real-life

multiple bearings.

2.2.6 Couple model


The purpose of the coupled model research is to explore a systematic framework and mathematical foundation to establish bonds between data-driven and physics-based models for Prognostics and Health Management (PHM) research. Existing PHM research efforts in the literature focus on either data-driven or physics-based methods, while both research directions may suffer from various difficulties and technical challenges in real-world applications. The proposed research aims to investigate advantages and limitations of both approaches to eventually establish a more effective and adaptive systematic PHM approach that closely integrates fundamental methodologies of both sides.


July 15, 2012 to July 14, 2014


1) To develop the methods of generalizing data driven models for the data sets from different conditions and develop a systematical method to select the best suitable model for the similar application cases in PHM.

2) To develop a systematical method to generate data from virtual physical models and establish the initial data-driven model by using the generated data set when sensor data is unavailable in real practical use or in the beginning phase.

3) To effectively accumulate and interactively integrate the knowledge from the physical model and driven model and establish the coupled PHM model.



PO Box 210072 F 1.513.556.3390


70


Figure 46. Research Plan for Coupled Model Scheme

The coupled model scheme (Figure 46) effectively combines the data driven-based prognostics and the physical based models to find not only the best suitable models among them, but also to develop more competitive integrated methods that can significantly reduce cost and increase the availability in PHM research. Previous efforts of using both data-driven and model based methods for PHM are very rare and application-specific. Comparing to traditional PHM approaches using either data-driven only or model based methods, the coupled model research utilizes advantages of both methods and is able to generate more accurate and stable prediction results. Case studies have shown that such approach provides a more optimized solution for solving PHM problems with high flexibility and the closely integrated workflow combining the best features from both approaches.

The developed coupled model has been validated in three different fields, including PHM for rotating machinery, prognostics for lithium-ion batteries in electric vehicles and adaptive control for high precision laser tracker systems. For the first case, previously developed physical models of bearings and ball screws are used to generate simulated training data. The training data is then used by machine learning algorithms SVM (Support Vector Machine) and SOM (Self-Organize Map) to build data-driven models for health assessment. In this case the physical models solve the common problem of limited availability of training data, especially for components that have many different specifications (e.g. geometric dimension, load types, etc.). The results of those case studies have been presented to member companies in the 23th and 26th IMS IAB meetings. For the third case, the coupled model approach is carried out through the self-tuning tracker project between IMS and a member company Automated Precision Inc. (API). Due to high precision requirement of the laser tracker system, a system dynamics model is built first. However, the physical model has difficulties to track variances between tracker



PO Box 210072 F 1.513.556.3390


71

systems and the degradation behavior over time. As a result, data-driven algorithms including least squares based system identification and prediction algorithms are added on top of the system dynamic model so that the tracker is able to automatically detect system deterioration and adjust control parameters accordingly. The system has proven to be effective and been tested for in-field applications by the member company.


The coupled model research and methodology have been successfully implemented and validated in several case studies. Moreover, further research has been done to analyze and extract common approaches and patterns of coupled model in a general form. Common approaches include typical algorithms and techniques from both data-driven and model based methods that can be used for integration at each stage of PHM, such as physical model based simulation for fast generation of training data, feature extraction based on machine failure expert knowledge and system identification methods using experiment data. The coupled model scheme is also integrated into Cyber-Physical System framework as the computational engine for accurate PHM in the cyber space. The research result has been presented during IAB presentations and published as journal and conference papers.



PO Box 210072 F 1.513.556.3390


72

2.3 PHM Society Data Challenges

2.3.1 [PHM 2008] A Similarity-based Prognostics Approach for Remaining

Useful Life Estimation of Engineered Systems


Each year the Prognostics and Health Management Society (PHM Society) sponsors a data challenge to both professional and student participants; participants use the available data and their prior experience and algorithm knowhow for developing their health monitoring algorithms for the given application. The accuracy of the health monitoring algorithm output is compared with the true health state; the true health state is unknown to the participants but is known by the organizers of the data challenge. The 2008 data challenge deals with jet engine remaining useful life prediction; the Center for Intelligent Maintenance Systems won the 1st and 3rd places of the competition, including both Professional and Students categories. This specific summary provides an overview of the winning technique designed by IMS researcher Tianyi Wang, which later was further developed into his Ph.D. dissertation.


May 2008 – August 2008


To develop accurate health prediction methods, with given full lifecycle datasets as training input.

To benchmark with other competing university and industry methods.

To document and further improve the methods.


The simulated data set consists of multivariate time series that are collected from multiple. Each time series is from a different instance of the same complex engineered system. There are three operational settings that have a substantial effect on unit performance. The data for each cycle of each unit include the unit ID, cycle index, 3 values for the operational settings and 21 values for 21 sensor measurements. The sensor data are contaminated with noise. Each unit starts with different degrees of initial degradation and manufacturing variation that is unknown. The unit is operating normally at the start of each time series, and develops a fault at some point during the series. The data set is further divided into training and testing subsets. In the training data set (218 units), the fault grows in magnitude until system failure. There is no "hard" failure in the data set; however, the remaining useful life of the last operational cycle of each unit in the training data is considered as zero. In the testing data set, the time series ends some time prior to system failure. The objective of the problem is to predict the number of remaining operational cycles before failure in the testing data set. A portion of the testing data set (218 units) is provided first to assist algorithm development and the rest (435 units) is released towards the end of the competition as the validation data set to score the algorithm.



PO Box 210072 F 1.513.556.3390


73

The algorithm flowchart is shown in Figure 47, where seven steps are divided into two stages – training (model development) and testing (RUL estimation).

Figure 47. Flowchart of the SBTP algorithm for jet engine remaining useful life prediction

During training stage, the training dataset is first partitioned into different operating regimes based on the operation settings. Sensor selection is achieved based on the trend of its reading during the unit’s lifetime. Selected sensor readings are fed into performance assessment modeling, where the model is obtained by fitting the sensor reading with parametric curve models. The best model is then selected during Model Identification step.

During testing stage, the first three steps in training stage are utilized first to classify the testing data of each testing unit. Then, the testing data is filtered with moving average method, compared with each of training model with a certain distance metric (e.g. Euclidean Distance), and the prediction result is obtained by fusing all the comparisons.

e) Project Deliverable:

The design and configuration of an innovative prediction algorithm, based on abundant training life cycles.

Publications (both conference and journal)



PO Box 210072 F 1.513.556.3390


74

2.3.2 [PHM 2009] Gearbox Health Assessment and Fault Classification


Each year the Prognostics and Health Management Society (PHM Society) sponsors a data challenge to both professional and student participants; participants use the available data and their prior experience and algorithm knowhow for developing their health monitoring algorithms for the given application. The accuracy of the health monitoring algorithm output is compared with the true health state; the true health state is unknown to the participants but is known by the organizers of the data challenge. The 2009 data challenge application deals with gearbox health monitoring and diagnosis; the Center for Intelligent Maintenance Systems had the top 3 teams for the 2009 competition. This summary contains an overview of the method and algorithm used by the team that finished second overall and first among all student competitors.


June 2009 – September 2009


To develop accurate health monitoring and diagnostic methods for gearbox components To benchmark the Center for Intelligent Maintenance Systems methods with other

competing university and industry methods. To document the method and results


As shown in Figure 1, the health monitoring method consisted of a set of algorithmic processing step. These steps included extracting features from the vibration signal, segmenting the data by operating regime, an algorithm for determining the overall health condition of the gearbox, and diagnosing which components in the gearbox are in a degraded condition. Figure 48 also contains the method highlights, which include the use of multiple feature extraction tools for detecting the different type of bearing and gear faults, a regime identification algorithm, a health assessment method that correctly identified all the baseline files, and a modified distance method for fault identification.

Some examples of the signal processing and feature extraction methods used in this application are provided in Figure 49. Kurtosis is a measure of the impulsiveness in the signal, a degraded bearing or gear wheel component in a gearbox would typically exhibit an impulsive signature. This particular kurtosis indicator is plotted as a function of frequency in the left-most plot in Figure 2 and can clearly discriminate between a healthy and degraded gearbox. The right-most plot in Figure 49 highlights the use of wavelets for detecting transient gear tooth related faults; again the impulsive signature from a broken gear tooth wheel is quite noticeable. Figure 50 provides another example of one of the detected fault cases from the algorithm. The left-most plot shows the signature of the bent shaft and the right-most plot shows the algorithm output and the correctly identified files (4 of them) that had this fault condition.



PO Box 210072 F 1.513.556.3390


75

Figure 48. Overview of gearbox health monitoring method

Figure 49. Example signal processing methods

Figure 50. Bent shaft signature and probability of fault result

Flow Chart of Systematic Method

Method Highlights

Using several signal processing

techniques provided the best feature set

Able to segment data by gear type,

load , and speed.

Used modified distance calculation

for vibration based fault

identification

Correctly identified all 80 healthy

baseline data sets using health

assessment method



PO Box 210072 F 1.513.556.3390


76


Development and refinement of IMS tools for gearbox health monitoring Journal paper and presentation material that document the method and results

2.3.3 [PHM 2010] Remaining Useful Life Estimation of High-speed CNC Milling

Machine Cutters


The 2010 challenge was focused on RUL estimation for a high-speed CNC milling machine cutters using dynamometer and accelerometer data.

Tool wear monitoring and predicting the RUL of the cutters have been a research area of interest since the wear in a milling cutter lowers the surface quality of the product and its dimensional accuracy. Avoiding this can improve the quality of finished parts and make this process more efficient.

The provided data consisted of collected force and acceleration signals in three directions along with the acoustic emission signals. The data was provided from six cutters and the participants were expected to estimate the remaining useful life of the cutters as they cut the work-piece. The actual wear measurement for three of the cutters was already provided along with the data to be used for training the models.

This method ranked 3rd in the student category of the competition.


June 2010 – August 2010


To develop a prognostics algorithm for estimating the remaining useful life of the milling machines


The technical approach consisted of four stages which are explained below:

De-noising the signals: The wavelet de-noising technique was used to remove the unwanted components from the signals. The mother wavelet Db2 at level 3 was used.

Extracting the features from the signals: The features were extracted by calculating the standard deviation of the wavelet coefficients from the vibration and force signals. The mother wavelet Db3 at level 5 was used.

Thirty six features were extracted out of all the six signals for each cutter. Using the correlation coefficient, the features that had the highest correlation with the output were



PO Box 210072 F 1.513.556.3390


77

picked up. As can be seen from the figure next page, 33 of 36 were chosen to build up the feature vector.

Modeling the wear: for modeling the wear of the cutters, the multi-layer perceptron neural network was implemented. This method basically connects the inputs (extracted features) to the output (measured wear) via units called neuron. Each connection has a weight value and transfer function. The weight values change during each cycle to find the best correlation between the inputs and the output. After optimizing the weight values, they are set and then the test data feeds into the network to provide us the output, which is the wear in this application.

RUL estimation using the modeled wear: This stage was done by calculating the maximum number of cuts for a given wear limit of 66x10-3 to 165x10-3 mm as it was instructed by the competition organizers.

Figure 51 shows an overview of the general approach.


Development of a prognostics technique for RUL estimation of the milling machine cutters

Documentation of the method for further applications



PO Box 210072 F 1.513.556.3390


78

Figure 51. The general procedure for RUL estimation of the milling machine cutters

a) De-noising

b) Wavelet Decomposition

c) Feature Extraction

d) Feature Selection e) Neural Network



PO Box 210072 F 1.513.556.3390


79

2.3.4 [PHM 2011] Fault Detection of Anemometers


Each year the Prognostics and Health Management Society (PHM Society) sponsors a data challenge to both professional and student participants; participants use the available data and their prior experience and algorithm knowhow for developing their health monitoring algorithms for the given application. The accuracy of the health monitoring algorithm output is compared with the true health state; the true health state is unknown to the participants but is known by the organizers of the data challenge. The 2011 data challenge deals with anemometer sensor health for wind energy applications; the Center for Intelligent Maintenance Systems had the top overall score for the 2011 competition. This summary contains an overview of the method and algorithm used that provided the top result in the competition.




To develop accurate sensor health monitoring methods To benchmark with other competing university and industry methods. To document the method and results


Figure 52. Overview of anemometer health monitoring method



PO Box 210072 F 1.513.556.3390


80

The algorithm flow chart is shown in Figure 52 and consists of both a training and monitoring phase. Data pre-processing steps are done to remove outlier instances and also to normalize the wind speed measurements since the anemometers are at different elevations. A non-linear principal component analysis algorithm is used to learn the normal relationship between the sensor variables from the baseline data. After the model is trained, the wind speed measurements of the monitored system are input into the trained health monitoring model; a residual calculation is performed between the predicted and actual sensor measurements. An additional clustering step is used to enhance the detection capability since the sensor fault is intermittent and not present in all operating conditions.

An example of the predicted and actual sensor values for both a health system and one with a degraded sensor are provided in Figure 53. Notice that in the left-most plot, the predicted and actual wind speed values match up very closely. However, the right-most plot in Figure 2 is an illustration of a system with a faulty sensor in which the predicted and actual sensor values show some noticeable differences. Further examination of the residuals between the predicted and actual sensors showed a bi-modal distribution; Figure 54 provides a histogram of the residual distribution for the file with the degraded sensor. A health metric based on the mean of the most negative residual cluster provided a suitable metric for discriminating between data files from a health system and ones with degraded anemometer sensors.

Figure 53. Example residual processing method – Blue (Measured), Black (Predicted)

Figure 54. Histogram of residual sensor value and clustering result



PO Box 210072 F 1.513.556.3390


81

d) Project Deliverables:

Development and refinement of IMS tools for sensor health monitoring Journal paper and presentation material that document the method and results

2.3.5 (2014) An Effective Predictive Maintenance Approach based on

Historical Maintenance Data using a Probabilistic Risk Assessment


Each year the Prognostics and Health Management Society (PHM Society) sponsors a data challenge. The 2014 data challenge deals with failure prediction of unknown industry assets provide by GE; the Center for Intelligent Maintenance Systems had the top overall score for the 2014 competition. This specific summary provides an overview of the winning technique designed by IMS researcher Seyed Mohammad Rezvanizaninai, which later was further developed into his Ph.D. dissertation.

The objective of the data challenge is to design an algorithm that would evaluate whether or not a system would fail in the next three days after a given sample time. The dataset consisted of three training datasets and two test datasets. The training data sets included part consumption, usage and failures. The test data sets included part consumption and usage. The part consumption data contained a record of what parts were replaced on an asset, the time that the part were replaced, the reason the asset was being worked on, and the quantity of parts that were replaced.




To develop a prediction technique for given sample time To estimate the high risk or low risk of the given time to adjust the maintenance activities To document the method and results


Figure 55 shows a sample of maintenance tasks for a sample asset. The 'x' axis shows the time in days for three years.



PO Box 210072 F 1.513.556.3390


82

Figure 55. Data description

The blue dashed line separates the testing part from the training part at the end of the second year. The 'y' axis shows the repeated number of part numbers replaced (RNPNR) in the maintenance time period with green color indicating replacement during training time and yellow color indicating replacement during testing time. The RNPNR means how many times part numbers have been repeated at each maintenance time in the part consumption regardless of the unique number of part numbers.

Figure 56. a) General Bathtub curve of industrial machines b) PM affect on decreasing hazard

rate

Dealing with complex reparable systems, PMs usually are executed on the wear out section to reduce hazard rate. Such PM operations when performed bring the system to operate close to new state. After each maintenance activity there is a short period of high hazard rate same as



PO Box 210072 F 1.513.556.3390


83

the infant mortality in the Figure 56 a) and b). These maintenance actions can be preventive maintenance with large number of parts replaced or even corrective maintenance (CM) with a few parts replacement. The main difference is that for PMs the higher failure rate is before the PM occurs, however, for CMs it is after that. The proposed methodology uses both techniques and builds a model that infers high risk before the PMs and after CMs.

Figure 57.Flow chart of "PM detection algorithm"

Pre-Processing functions, included data observation, sorting and irregularities in the usage data, were used to eliminate data that does not have the quality required to accurately determine features. The concept of a bathtub curve was used to extract and select features on how soon after a maintenance a failure would occur during the training data, which was then applied to the testing data to provide an accurate health assessment rating and therefore determine if the asset was at high risk. A crucial part of using the bathtub curve is determining where the preventative maintenances fall in the data we were given. Through observation and analysis of the repeated number of part numbers replaced, and MTBR, function were used to determine where the preventative maintenances fell (Figure 57), which provided most of the information needed to determine the heath risk of assets in the given data set.



PO Box 210072 F 1.513.556.3390


84

Figure 58. Histogram of failure after CMs

The other approach is to use the corrective maintenances to evaluate the correlation between corrective measures and the probability of failure afterword (Figure 58). Using the number of failures immediately after maintenance actions in the testing data, the probability of failure after maintenance actions was calculated.

e) Project Deliverables:

Development risk probability technique for failure prediction Journal paper and presentation material that document the method and results



PO Box 210072 F 1.513.556.3390


85

2.4 Biography: Professor Jay Lee

Dr. Jay Lee is Ohio Eminent Scholar and L.W. Scott Alter Chair Professor in Advanced Manufacturing at the University of Cincinnati. Previously, he held a position as Wisconsin Distinguished Professor and Rockwell Automation Professor at the Univ. of Wisconsin-Milwaukee and is founding director of National Science Foundation (NSF) Industry/University Cooperative Research Center (I/UCRC) on Intelligent Maintenance Systems (IMS www.imscenter.net ) which is a multi-campus NSF Center of Excellence between the Univ. of Cincinnati (lead institution), the Univ. of Michigan, and the Univ. of Missouri-Rolla. Prior to joining UWM, he served as R&D Director for Product Development and Manufacturing Department at United Technologies Research Center (UTRC), E. Hartford, CT, and was responsible for the strategic direction and R&D activities for next-generation products and manufacturing, and service technologies. Prior to joining UTRC, he served as Program Directors for a number of programs at NSF during 1991-1998, including the Engineering Research Centers (ERCs) Program, the Industry/University Cooperative Research Centers (I/UCRCs) Program, and the Materials Processing and Manufacturing Program (MPM). In addition, he had served as an adjunct professor for a number of academic institutions, including Johns Hopkins University, where he was an adjunct faculty member for the School of Engineering and Applied Science as well as for the Hopkins Technical Management Program during 1992-1998. He conducted research work at the Mechanical Engineering Lab. of the Ministry of International Trades and Industry (MITI) as a Japan Science and Technology Agency (STA) Fellow in 1995, a Japan Society for Promotion of Science (JSPS) Fellow at the Univ. of Tokyo as in 1997, and a visiting professor at Swiss Institute of Technology (EFFL), Lausanne, Switzerland in July 2004.. His current research focuses on autonomous computing and smart prognostics technologies including predictive machine degradation assessment, remote monitoring, embedded prognostics, and self-maintenance systems. He had served on Board on Manufacturing and Engineering Design (BMAED) of National Research Council, Board of Directors for the National Center for Manufacturing Science (NCMS), Chairman of the Manufacturing Engineering Div. and Materials Handling Engineering Div. of ASME, etc. He has authored/co-authored over 100 technical publications, edited two books, contributed numerous book chapters, three U.S. patents, 2 trademarks, and had delivered numerous invited lectures and speeches, including over 100 invited keynote and plenary speeches at major international conferences. Dr. Lee received his B.S degree from Taiwan, a M.S. in Mechanical Engineering from the Univ. of Wisconsin-Madison, a M.S. in Industrial Management from the State Univ. of New York at Stony Brook, and D.Sc. in Mechanical Engineering from the George Washington University. He received Milwaukee Mayor Technology Award in 2003 and was a recipient of SME Outstanding Young Manufacturing Engineering Award in 1992. He is also a Fellow of ASME and SME.

http://www.imscenter.net/

Documents

IMS Project Portfolio - 2015