detection, isolation, and recovery of sensor...
TRANSCRIPT
Detection, Isolation, and Recovery Detection, Isolation, and Recovery of Sensor Performance Degradationof Sensor Performance Degradation
Research Assistants/Staff: Faculty:L. JIANGL. JIANG D. D. DjurdjanovicDjurdjanovic, , JJ.. NiNi
For more information, contact Prof. J. Ni; Phone: 734-936-2918; Email: [email protected]
ObjectivesObjectives
StateState--ofof--thethe--ArtArt
Electronic Automotive Throttle SystemElectronic Automotive Throttle System
Future WorkFuture Work
• An ever-increasing number of sensors have been employed in dynamic systems to ensure correct control functionalities and diagnosis.
• Nevertheless, a sensor itself degrades and fails, just as any other dynamic system.
• A faulty sensor may cause undesirable system performance, process shut down, or even a fatal accident.
Instrument Fault Detection and Isolation Techniques
Hardware Redundancy
Analytical Redundancy
Model-based method
Knowledge-based expert system
Process history based method
Parity relations
Luenberger observer and Kalman filter
Parameter estimation
Multivariate statistical methods
Bayesian belief networks
Artificial neural networks
Desai and Ray (1984) Glockler et. al (1989)
Qin and Li (1999, 2001)
Van Dirk (1994) Upadhaya and Kerlin (1978)
Lee (1994) Betta et al. (1992)
Dunia et al., 1996 Ibarguengoytia et al. (2006) Mehranbod et al. (2003)
Betta et al., (1998) Napolitaneo et al. (1994)
ApproachesApproaches
Fit a proper model to the data within the moving window by subspace model based method
x(t+Ts) = A(t)x(t)+B(t)y(t)+K(t)e(t)y(t) = C(t)x(t)+D(t)y(t)+e(t)
Extract the time constants of the monitored system and the sensor from the poles of the
corresponding transfer function
Degrading Monitored System
Normal Compound System
Changes in slow dyanamics?
Y
Degrading Sensor
Changes in fast dyanamics?
Y
ANDN N
T
Assumption 1: There is no nonlinearity included in the system
Assumption 2: Sensor dynamics is much faster than that of the plant
Detection and Isolation of Dynamic ChangesDetection and Isolation of Dynamic Changes
Detection and Isolation of Gain ChangesDetection and Isolation of Gain Changes
Assumptions 3: The process disturbances and measurement noise are white noise
Monitored Plant Gp(s) Sensor Gs(s)
Control Signals u(t)
Sensor Measurements y(t)
Process Disturbances wp(t)
Measurement Noise wn(t)
++ +
+
( ) ( ( ) ( )) ( )p s p s nY s G G U s W s G W s= + +
Confidence Value
10 20 30 40 50 60 70 80 90 100
0.2
0.4
0.6
0.8
1
10 20 30 40 50 60 70 80 90 100
0.2
0.4
0.6
0.8
1
Time (sec)
Time Constants
0 10 20 30 40 50 60 70 80 90 1000.16
0.18
0.2
0.22
0.24
0 10 20 30 40 50 60 70 80 90 1000.02
0.03
0.04
0.05
Time (sec)
Sen
sor
Pla
nt
0 10 20 30 40 50 60 70 80 90 100
0.8
1
1.2
1.4
p
0 10 20 30 40 50 60 70 80 90 1000.4
0.6
0.8
1
1.2
1.4
Normalized Gains
Time (sec)
Sens
orPl
ant
10 20 30 40 50 60 70 80 90 1000
0.2
0.4
0.6
0.8
1
10 20 30 40 50 60 70 80 90 1000
0.2
0.4
0.6
0.8
1
Confidence Value
Time (sec)
• Develop methods for the case when the process disturbances and measurement noise are colored noise processes.
• Extend this method to detect and isolation degraded sensors in presence of nonlinearities in the monitored system
A Watchdog AgentA Watchdog Agent™™ Toolbox for Toolbox for MultisensorMultisensor Performance AssessmentPerformance Assessment
Research Assistants/Staff: Faculty:K. JohnsonK. Johnson D. D. DjurdjanovicDjurdjanovic,, J. NiJ. Ni
For more information, contact Prof. J. Ni; Phone: 734-936-2918; Email: [email protected]
ObjectivesObjectives
StateState--ofof--thethe--ArtArt
ApproachesApproaches
AccomplishmentsAccomplishments
Future WorkFuture WorkSponsorsSponsors
• The Watchdog Agent™ Toolbox is a software program that has been created to assess and predict system performance in various conditions and easily assess which combination of algorithms is the best fit for a particular application.
• Several interchangeable tools have been developed in Matlab and compiled into a standalone application. The tools that have been integrated can be seen below in the main window of the Watchdog Agent™ Toolbox.
• The software takes as input sensor data from the current, normal, and faulty operation of the equipment, and outputs the current level of degradation of the system, an ARMA prediction of future degradation, and a sensitivity analysis, ormeasure of affinity, for the chosen algorithms.
• The Watchdog Agent™ Toolbox has several functional layers, shown below, compliant with the Open System Architecture for Condition Based Maintenance (OSA-CBM).
• Modularization of the toolbox allows different methods of signal processing, feature extraction, performance evaluation, and sensor fusion to be used in any combination.
• A method for analysis of discrete event data will be added to the toolbox.
• NSF Industry/University Collaborative Research Center on Intelligent Maintenance Systems
• National Instruments
Data Acquisition National Instruments PXI system
Data Manipulation Time frequency, wavelet, FFT analysis
Condition Monitor Not currently integrated
Health Assessment
Statistical pattern recognition, logistic regression, sensor fusion
Prognostics ARMA modeling
Decision Support Not currently integrated
Presentation Graphical user interface (GUI) • Shown below is the degradation analysis from 1100 cycles of a boring tool at DaimlerChrysler; two methods of performance assessment are compared.
Statistical Pattern Recognition Logistic Regression• The following picture shows the National Instruments PXI
system which is used for data acquisition.
tool replacement tool
wear
High speed (up to 8 MHz) simultaneous sampling
PC based data acquisition
Fieldbus
Zone PLC
EthernetSwitch/Router
Ethernet
Switches Valves Pumps Motor DrivesRobot/MTControllers
“Polling”Status Monitoring
COSReporting
COSReporting
Embedded Watchdog AgentEmbedded Watchdog Agent™™ for for Industrial Automation SystemsIndustrial Automation Systems
Research Assistants/Staff: Faculty:Y. Lei Y. Lei D. D. DjudjanovicDjudjanovic, , J. NiJ. Ni
For more information, contact Prof. J. Ni; Phone: 734-936-2918; Email: [email protected]
ObjectivesObjectives
StateState--ofof--thethe--ArtArt
ApproachesApproaches
AccomplishmentsAccomplishments
Future WorkFuture Work
• Provide diagnostic & predictive information in a networked industrial automation system at the device level.
• Develop a methodology to evaluate effects of device-level embedded prognostics
• Create methods to minimize maintenance-related costs through strategic allocation of failure risks associated with basic embedded prognostics.
• Built two testbeds for verification and demonstration of effects and benefits of the basic prognostics in Embedded Watchdog Agent
• The Embedded Watchdog Agent™ solution
• The risk allocation problem
• Due to bandwidth limitations, limited amount of maintenance related information is available in an industrial networked automation systems.
• Inherent intelligence of networked devices is not utilized for maintenance purposes.
Level 1 & 2 EWA Testbed Sponsored By Rockwell Automation and GM
Smart DeviceNet Device Testbed Sponsored By Omron
PLC
Level 2 EWA I/O
Block
Cylinders
& Valve
PLC
Level 2 EWA I/O
Block
Cylinders
& Valve
Smart DeviceNet DeviceSmart DeviceNet Device
• Developed a risk allocation methodology for industrial automation systems with basic prognostics
340000
342000
3420
00
342000
3420
00
342000
342000
342000
344000
344000
3440
00
344000
344000
3440
00
344000
3440
00
3440
00
344000
344000
344000
344000
346000
346000
346000
3460
00
346000
346000
346000
346000
346000
3460
00
3480
00
348000
Failure Risk of Type 1 Machine (%)
Failu
re R
isk
of T
ype
2 M
achi
ne (%
)
0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
(0.0061, 0.0162) GA Searching Path
340000
342000
3420
00
342000
3420
00
342000
342000
342000
344000
344000
3440
00
344000
344000
3440
00
344000
3440
00
3440
00
344000
344000
344000
344000
346000
346000
346000
3460
00
346000
346000
346000
346000
346000
3460
00
3480
00
348000
Failure Risk of Type 1 Machine (%)
Failu
re R
isk
of T
ype
2 M
achi
ne (%
)
0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
(0.0061, 0.0162) GA Searching Path
Risk Allocation of Risks of Failure Using Genetic Algorithm
• Developed Dynamic Risk allocation method using Basic Embedded Prognostics
• Conduct and improve dynamic optimal risk allocation using Basic Embedded Watchdog Agent™ Prognostics in complex industrial automation systems.
• Conduct industrial test and cost effect analysis of basic prognostics Embedded Watchdog Agent™.
• Defined 3 levels of EWA complexity
Flowchart for optimal risk allocation based on a Genetic Algorithm
Machine
Buffer
Machine
Buffer
??
???
?
??
OptimizationOptimization
Level 1:Operation/cycle counter
Level 2: Operation/cycle counter AND Cycle Timer
Level 3: Trending and data streaming
N N+1
Time
1 (on)
0 (off)
N N+1
Time
N N+1
Time
1 (on)
0 (off)
N N+1
Time
ΔT ΔTN N+11 (on)
0 (off)
N N+1
Time
ΔT ΔTN N+11 (on)
0 (off)
FeatureExtraction
Multisensor Perf. Evaluation
Performance Assessment
Learning
Performance Prediction
SensorySig. Proc.
Condition Diagnosis
FeatureExtraction
FeatureExtraction
Multisensor Perf. Evaluation
Multisensor Perf. Evaluation
Performance Assessment
Learning
Performance Prediction
SensorySig. Proc.Sensory
Sig. Proc.
Condition Diagnosis
Basic P
rognostics
Advanced Prognostics
Basic P
rognostics
Advanced Prognostics
Table: Risk levels assigned to machines of Type 1 & Type 2
layout of the system considered
NetworkFeature
Extraction
Performance Assessment
Performance Prediction
Failure modeDatabase
Watchdog Agent
NetworkNetworkFeature
ExtractionFeature
Extraction
Performance Assessment
Performance Assessment
Performance Prediction
Performance Prediction
Failure modeDatabase
Failure modeDatabase
Watchdog Agent
Network Watchdog AgentNetwork Watchdog Agent™™ for for Industrial Automation SystemsIndustrial Automation Systems
Research Assistants/Staff: Faculty:Y. Lei Y. Lei D. D. DjudjanovicDjudjanovic, , J. NiJ. Ni
For more information, contact Prof. J. Ni; Phone: 734-936-2918; Email: [email protected]
ObjectivesObjectives
MotivationMotivation
ApproachesApproaches
AccomplishmentsAccomplishments
Future WorkFuture Work
• Develop and evaluate novel network health monitoring tools for industrial automation systems
• Develop methodology that could assess and predict the performance of industrial automation networks.
• Identified key failure modes that occur in the networked industrial automation systems
• Proposed data acquisition system
• Nearly 54% of failures in a networked industrial automation system occur due to faults on the network itself.
• There is need to create a predictive agent for assessment and prediction of performance of an industrial automation network.
• Current tools are not designed to provide such capabilities
• Develop NWA for the data link layer
• Develop diagnosis methodology using data from both data link layer and physical layer
• Testbed validation and demonstration for NWA.
Measurement of key parameters of DeviceNet
PC
DeviceNet Interface Card
Data Acquisition Card
Statistic DataError Frame Ratio
Bus loadBus VoltageOvershoot
Signal-To-Noise Ratio
Framework of the Network Watchdog Agent™
• Connection problems, including loose connection of cable connectors and terminating resistors• Corrupt frames, including damaged transceivers, grounding problems, electrical Magnetic Interference (EMI)
• Identified key parameters that affect the performance of networked industrial automation systems
• Bus voltage and current• Common Mode Signal features• Waveform of Digital signal
• Preliminary results of one lost terminating resistor
-2 0 2 4 6 8 10 12 14 16
x 10-6
1.5
2
2.5
3
3.5
4(a) Normal DeviceNet Signal Profile
Time(s)
Vol
taag
e(V
)
CANLCANH
-2 0 2 4 6 8 10 12 14 16
x 10-6
1
1.5
2
2.5
3
3.5
4
(b) DeviceNet Signal Profile When Loosing One Terminating Resistor
Time(s)
Vol
tage
(V)
CANLCANH
(b) DeviceNet Signals When Loosing One Terminating Resistor
Time (s)
Time (s)
(a) Normal DeviceNet Signals
Vol
tage
(V)
Vol
tage
(V)
-2 0 2 4 6 8 10 12 14 16
x 10-6
1.5
2
2.5
3
3.5
4(a) Normal DeviceNet Signal Profile
Time(s)
Vol
taag
e(V
)
CANLCANH
-2 0 2 4 6 8 10 12 14 16
x 10-6
1
1.5
2
2.5
3
3.5
4
(b) DeviceNet Signal Profile When Loosing One Terminating Resistor
Time(s)
Vol
tage
(V)
CANLCANH
(b) DeviceNet Signals When Loosing One Terminating Resistor
Time (s)
Time (s)
(a) Normal DeviceNet Signals
Vol
tage
(V)
Vol
tage
(V)
Comparison of DeviceNet signals
under normal condition and with
one lost terminating resistor
Feature cluster under normal condition and
with one lost terminating resistor
Data collection system
• National instruments digital oscilloscope
• SST DeviceNet interface card
0 100 200 300 400 500 600 7000
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
conf
iden
ce v
alue
time (second)
Plant AD
confidence valuecontrol limit
0 100 200 300 400 500 600 7000
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
conf
iden
ce v
alue
time (second)
Controller AD
confidence valuecontrol limit
Anomaly Detection and Localization Anomaly Detection and Localization for Dynamic Control Systemfor Dynamic Control System
Research Assistants/Staff: Faculty:JJ. . LiuLiu J. NiJ. Ni
For more information, contact Prof. J. Ni; Phone: 734-936-2918; Email: [email protected]
ObjectivesObjectives
StateState--ofof--thethe--ArtArt
ApproachesApproaches
AccomplishmentsAccomplishments
Future WorkFuture Work
SponsorsSponsors
• Detecting and localizing unknown faults i.e. anomalies
• Sensing gradual performance degradation of controlled dynamic system
• Discriminating whether the performance deviation is caused by control software or plant (hardware)
• Developing effective diagnostics that can be applied comprehensively from product design through design validation, manufacturing testing on-board monitoring and finally repair service operating
• Ability to detect gradual changes of the system parameters as the dynamic system operates.
• Ability to detect faults that were not observed before (i.e. detect anomalies) using the same generic approach.
• No prior assumptions were made about the system behavior and the model of system behavior was learnt from a normal driving profile.
• Ability to detect and isolate of controller (software) anomalies from the plant (hardware) anomalies.
• With the growing complexity of both plants and control systems, effective diagnostics for all possible failures has become increasingly difficult and time consuming.
• As a result, many systems rely on limited diagnostic coverage provided by a diagnostic strategy which test only for known or anticipated failures, and presumes the system is operating normally if the full set of tests is passed.
• In addition, these tests are often developed separately and at large costs in terms of time and resources after hardware and the control system have been produced and are available for analysis.
• Development of effective method for intermittent fault detection
Inputs
Initial Conditions
OutputsDynamic System
Region of Operation
SOMs TFA & PCA
Performance Index
• The new anomaly detection method relies on the use of Self-Organizing Maps (SOM) for regionalization of the system operating conditions, followed by Time-Frequency Analysis (TFA) and Principal Component Analysis (PCA) for anomaly detection and fault isolation
• The proposed approach consists of two elements: anomaly detection which identifies whether a deviation from normal operation has occurred and fault isolation, which isolates the problem, as best as possible, to the specific component or subsystem that has failed.
• This project is supported in part by Center for Intelligent Maintenance System at University of Michigan, Ann Arbor and ETAS Inc., Ann Arbor
Method Overview
Anomaly Detection and Localization
Driving Profile:FTP75
AD on controller AD on plant
Gradual decrease in the gain factor (parameter of controller)
0 500 1000 1500 2000 2500 3000 3500 40000
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
conf
iden
ce v
alue
time (seconds)
cv from detector for F1 cv from detector for F2cv from detector for F0control limit
Fault isolation
0 500 1000 1500 2000 2500 3000 3500 40000
0.5
1
1.5
2
2.5
3
time (seconds)
Inde
x
F0
F1
F3
F2
F0
Index Category
F0 Normal
F1 Reduced K
F2 Increased C
F3 A unknown fault
Data Fusion and Predictive Modeling for Data Fusion and Predictive Modeling for Intelligent Maintenance in Complex Intelligent Maintenance in Complex
Semiconductor Manufacturing ProcessesSemiconductor Manufacturing Processes
Research Assistants/Staff: Faculty:YY. . Liu Liu J. Ni, D. DjurdjanovicJ. Ni, D. Djurdjanovic
For more information, contact Prof. J. Ni; Phone: 734-936-2918; Email: [email protected]
ObjectivesObjectives
StateState--ofof--thethe--ArtArt
ApproachesApproaches
AccomplishmentsAccomplishments
Future WorkFuture Work
SponsorsSponsors
• To develop system modeling and decision support tools that integrate different data domains in a semiconductor fabrication facility into coherent performance information that will facilitate rapid and proactive countermeasures for any problem that will adversely affect the yield and uptime
• To coordinate and synthesize the integrated information from all stations into system level information that will be used to make dynamic, accurate and cost-effective maintenance scheduling and product routing decisions
• Framework for data integration and probability inference
Process
Action
Inspection
Inspection
Action
Detection
Sensor
Action
Diagnostics
Historical DB
Action
Reliability Info
ProcessInspection
QualityInspection
EquipmentMonitoring Reliability
Uncoordinated& ReactiveActions
Uncoordinated& ReactiveActions
Prognostics Tools Decision Support Tools
Monitor > Correlation > Prediction > Prevention > Optimization
P(Yield| Quality, Sensing, Reliability, Process) Decision Tools, Optimization, Synchronization
-Coordinated& ProactiveActions-Rapid Response-High Yield-Zero Downtime-Rapid Cycles
-Coordinated& ProactiveActions-Rapid Response-High Yield-Zero Downtime-Rapid Cycles
Now
Prop
osed
Paradigm Shift –High Yield Next Generation Fab
Historical DB
Action
Maint. Info
Maintenance
• Many yield models based on defect detection and critical area analysis have been developed, in which, however, only defect inspection data are utilized
• Researchers have attempted to predict yield using localized in-line data or in-situ measurements of particle contamination, which only give a partial picture of the process degradation
• Majority of maintenance operations are still based on either historical reliability of equipment, or diagnostic information from equipment performance signatures extracted from in-situ sensors
• Improved yield prediction from integrated information flow using enhanced Bayesian networks
• Enhanced maintenance decision support tools based on improved yield estimation model
• Validation of results in an industrial testbed
Semiconductor Research Corporation (SRC)
• Performing case study using semiconductor dataset
• Handling incomplete data without compromising inference accuracy
• Periodically updating inference structure when real inspection becomes available
• Simulation study to validate the proposed method
• Application to automotive industry dataset
Embedded Life Cycle Unit for Embedded Life Cycle Unit for Prognostics of Mechanical Prognostics of Mechanical
SystemsSystemsResearch Assistants/Staff: Faculty:R. R. Muminovic Muminovic Dr. Dr. D. D. Djurdjanovic Djurdjanovic / Prof. J. Ni/ Prof. J. Ni
ObjectivesObjectives
Condition Based MaintenanceCondition Based Maintenance
• Shift the maintenance policy from reactive to proactive• Assess current performance of a railroad bogie bearing
Sensors
Maintenance
Central Control Station
Signal Pre-Processingand Analysis (LCU & Diagnostic Algorithms)
PerformanceAssessmentand Prediction
ApproachApproach
Diagnostic and Prognostic Algorithms
Life Cycle unit and diagnostic/prognostic algorithms
Condition based maintenance policy
Feature extraction – Time-frequency distribution
• Time-frequency representation is a view of a signal represented over both time and frequency.
Normalized time
Nor
mal
ized
freq
uenc
y
• Time-frequency analysis is particularly suitable for nonstationary signals whose frequency content varies over time
• High identification rate of functional and defective bearing
• Improved maintenance policy through condition based maintenance
Rifet Muminovic <[email protected]>Prof. Jun Ni <[email protected]>Dr. Dragan Djurdjanovic <[email protected]>Daniel Odry <[email protected]>
Measured vibrations
Measurement: Measured vibration and given RPM
• RPM was varied with a sinusoidal input
• Sinusoid has a period of T=60s
• Experiment was conducted with five different loads acting on the bogie cage
Support Vector Machine (SVM)Input Space Feature Space• Training vectors are
mapped into a higher dimensional space with help of a kernel
• SVM algorithm creates a maximum-margin hyperplane that separates data into two classes
Testbed
* Mapping between input and feature space
Experimental setup
MeasurementMeasurement
Expected ResultsExpected Results
For more information, contact:For more information, contact:
Classification MethodClassification Method
Immune Systems Engineering Immune Systems Engineering for Complex Dynamic Systemsfor Complex Dynamic Systems
Research Assistants/Staff: Faculty:R. Torres R. Torres J. NiJ. NiD. DjurdjanovicD. Djurdjanovic
For more information, contact Prof. J. Ni; Phone: 734-936-2918; Email: [email protected]
ObjectivesObjectives
StateState--ofof--thethe--ArtArt
ApproachApproach
AccomplishmentsAccomplishments
Future WorkFuture Work
SponsorsSponsors
• Develop novel approach to realize “immune systems”functionalities in automated systems
• Robustly detect abnormal behaviors, isolate its source, and compensate for negative effects to achieve desired performance in spite of the presence of an anomaly
• Apply local adaptive controller to the local models
• Vary parameters to simulate anomalies
• Achieve desired performance despite anomalies
• Apply same approach to practical data (i.e. automotive engine)
National Science Foundation
• Nonlinear plant is regionalized into several local linear models through “divide and conquer” methodology
• Local models are analyzed for fault detection and isolation
• Adaptive local controllers are applied per region to achieve desired performance
Benchmark Problem• Continuously Stir Tank Reactor
(CSTR) Model is used to serve as the plant for preliminary study
• Inputs (flow rate, qr and heat transfer removal rate, QC) are used to control output (concentration Cb)
Model using Growing Structure Multiple Model System (GSMMS) algorithm
Reconfigure control system to restore desired performance
Identify anomalies via Anomaly Detection Agents
Diagnose and isolate faults
Are modeling errors large?
Continue control with current GSMMS model
YESNO
Local Linear Models of Nonlinear SystemOperational Space
Local Model
w1
1 1 1ˆ= −e y y
Adaptive Local Controller
(C1)
Local Model(A1)
1y
Updating SchemeVia AD1
NominalController
v1
∑
ref
qr
Cb
QC
CSTR Schematic
Modeling Results
• Recursive Learning for both GSMMS and Fuzzy Logic have been performed on CSTR model
• The index, variance accounted for (VAF), has been used to quantify the deviation of the models from the actual plant
(Note: the VAF of two equal signals is 100%)
Comparison of ResultsGSMMS
Computation Time = 2.01 sec
VAF = 99.1833 %
FuzzyComputation Time = 29.87 sec
VAF = 99.7115 %
50 55 60 65 70 75 80 85 90 95 1000.4
0.5
0.6
0.7
0.8
0.9
1
Time (min)
CB
CB vs Time
ActualGSMMSFuzzy
Results using GSMMS and Fuzzy