NSF I-UCRC on Intelligent Maintenance Systems;Planning Workshop at the University of Texas;May 15, 2012, Austin, TX
IMMUNITY INSPIRED DIAGNOSTICSAND PROGNOSTICS IN SYSTEMSOF INTERACTING DYNAMICSUBSYSTEMSPROF. DRAGAN DJURDJANOVIC
NSF I-UCRC on Intelligent Maintenance Sys.Planning Workshop at the University of Texas
2 PRESENTATION OUTLINE
»Motivation»Methodology»Preliminary Work»Research Plan
Time
Each dot is a FOUPof wafers
Particle failures – tooldown for several weeks
MOTIVATION:SEVERAL MONTHS OF PECVD TOOLOPERATION
3
Coulomb crystals
and particle failures
Again tool down forseveral weeks
5 FAULT ISOLATION VIA DISTRIBUTEDANOMALY DETECTION
ADA4
AD3
AD2AD1
ChargeAir Cooler
EGRValve
VNT-Turbine
EGRCooler
TCACα
IntakeManifold
Pm
ExhaustManifold
λNeng
Compressor
WHFMEngineControl
Unit (ECU)
ThrottleValve Engine
XVNT
WFUEL
XEGR
Xth
*ADA -AnomalyDetector
Neng
Block diagram of the mass airflow in a diesel engine providedby the Bosch Research Center, Pittsburgh
6
ADA4
AD3
AD2AD11
ChargeAir Cooler
EGRValve
VNT-Turbine
EGRCooler
TCACα
IntakeManifold
Pm
ExhaustManifold
λNeng
Compressor
WHFMEngineControl
Unit (ECU)
ThrottleValve Engine
XVNT
WFUEL
XEGR
Xth
Neng
AD13AD12
Negative selection in natural immunesystems (Dasgupta, 2006)
*ADA -AnomalyDetector
Block diagram of the mass airflow in a diesel engine providedby the Bosch Research Center, Pittsburgh
FAULT ISOLATION VIA DISTRIBUTEDANOMALY DETECTION
AnomalousBehavior
Compute CV
Currentbehavior
Normalbehavior
Plant
Operating Region
Localmodel 1
Localmodel 2
Localmodel 4
Localmodel 5
Localmodel 6
Localmodel 7
Piecewise Dyn. Model
+-
Anomaly Detection usingPiecewise Dynamic Models
8
J. Liu, D. Djurdjanovic, K. Marko and J. Ni, “Growing StructureMultiple Model System for Anomaly Detection and Fault Diagnosis”,Transactions of ASME, Journal of Dynamic Systems,Measurements and Control, Vol. 131, No. 5, pp. 051001-1 –051001-13, 2009
EGRCooler
Injected FuelEngine Speed
++
PI Controller−+
EGR Valve
φECULook-upTable
Exhaust Temp
Exhaust PressureIntake Man. Pressure
Ambient Temperature
Look-upTable
ADoverall
Precedent Free Fault Isolation in the Exhaust GasRecirculation (EGR) System of a Diesel Engine
TESISDYNAware, 2006
M. Cholette and D. Djurdjanovic, “Precedent-Free Fault Isolation in a Diesel EngineEGR System”, to appear in ASME J. of Dynamic Systems Measurements & Control2011
9
EGRCooler
Injected FuelEngine Speed
++
PI Controller−+
EGR Valve
φECULook-upTable
Exhaust Temp
Exhaust PressureIntake Man. Pressure
Ambient Temperature
Look-upTable
ADmassflow
ADcooler
TESISDYNAware, 2006
Precedent Free Fault Isolation in the Exhaust GasRecirculation (EGR) System of a Diesel Engine
10M. Cholette and D. Djurdjanovic, “Precedent-Free Fault Isolation in a Diesel EngineEGR System”, to appear in ASME J. of Dynamic Systems Measurements & Control2011
EGRCooler
Injected FuelEngine Speed
++
PI Controller−+
EGR Valve
φECULook-upTable
Exhaust Temp
Exhaust PressureIntake Man. Pressure
Ambient Temperature
Look-upTable
AD3
AD2 AD4 AD5
ADcooler
TESISDYNAware, 2006
Precedent Free Fault Isolation in the Exhaust GasRecirculation (EGR) System of a Diesel Engine
11M. Cholette and D. Djurdjanovic, “Precedent-Free Fault Isolation in a Diesel EngineEGR System”, to appear in ASME J. of Dynamic Systems Measurements & Control2011
EGRCooler
Injected FuelEngine Speed
++
PI Controller−+
EGR Valve
φECULook-upTable
Exhaust Temp
Exhaust PressureIntake Man. Pressure
Ambient Temperature
Look-upTable
TESISDYNAware, 2006
Anomalies in the EGR Valve(Clogging of the Valve)
12M. Cholette and D. Djurdjanovic, “Precedent-Free Fault Isolation in a Diesel EngineEGR System”, to appear in ASME J. of Dynamic Systems Measurements & Control2011
ADcooler
13
ADmassflow
AD4
AD2(Look up table)
AD3(Look up table)
AD4(PI Controller)
AD5(Valve)
Detection and Isolation of an Anomaly in the EGR Valve
15 TASK 1: DISTRIBUTED ANOMALYDETECTION WHEN CAUSAL CONNECTIONSARE NOT CLEAR
• Physics basedreasoning and expertknowledge• Bayesian causationdiscovery approaches
NSF I-UCRC on Intelligent Maintenance Sys. Planning Workshop at the University of Texas
17 FAULT ISOLATION VIA DISTRIBUTED ANOMALY DETECTI ON
ADA 4
NSF I-UCRC on Intelligent Maintenance Sys. Planning Workshop at the University of Texas
16 TASK 2: POLICY FOR DISTRIBUTING ANOMALYDETECTORS IN A COMPLEX SYSTEM
16
d0 – Detection ratefor groupedanomaly detector
d1 – Detection ratefor individual FRUanomaly detectors CC – Computational cost per
unit time Ci – Cost of failing to localizethe fault per unit time
17 TASK 3: SIGNATURE PREDICTION FORMAINTENANCE DECISION-MAKING
• Existing methods are eitherdealing with stationary time-seriesor are too slow to provide real-timeinformation• Combine analytics withsimulation based tools to obtainaccurate predictions fast
NSF I-UCRC on Intelligent Maintenance Sys. Planning Workshop at the University of Texas
20 TASK 1: UNDERSTANDING UNCERTAINTY OF HIDDEN MARK OV MODEL ESTIM ATIO N
Similarity basedprediction
methods forpredictingsignature
distributions
Warning about adegraded state
Root cause of thedegraded state
18
POTENTIAL APPLICATION AREAS
• Automotive engines• Aircraft jet engines• Advanced
manufacturingequipment
• Electrical generatorsystems
• Modular data centers
SUMMARY OF PROPOSED WORK19
Task 1: Identification of causalrelations between subsystems in acomplex system
Task 2: Optimization of policies fordistribution of anomaly detectorsthat achieve tradeoff betweenspeed and accuracy of faultlocalization and computation load
Task 3: Implementation ofdegradation monitoring techniquesto a real system
• Mentors: Prof. DraganDjurdjanovic & industrialpartners
• Researchers: One graduatestudent per application areain year 1 and 2 graduatestudents in year 2 (possibleapplication areas includemanufacturing equipment,engines, electricalgenerators)
YEAR 1
1 2 3 4 5 6 7 8 9 10 11 12 1
TASKS
1
2
3
Ability to deal with unprecedented faults(faults never seen in the past)Data mining to understand relationsbetween subsystems of a highly complex(multi-physics) systemPotentials to dramatically shortenequipment downtimes and their impacts
• Method for identifying relationsbetween subsystems in a highlycomplex system of interactingdynamic subsystems
• Demonstration of faultlocalization in a complexsystem in which causalrelations between subsystemsare not a priori known(semicond. dataset)
20 BUDGET SUMMARY
• Development of the method forcausation discovery and oneapplication area, such as plasmabased semiconductormanufacturing tool
• $40K (student tuition andstipend for a year)
• Budget for a second applicationexample
• Additional $20K (student tuitionand support for 6 months)
• Total: $40-60K
Year 1 Budget Year 2 Budget
• Development of the method foroptimization of policies for distributionof anomaly detectors, with oneapplication area, such as plasma basedsemiconductor manufacturing tool orautomotive engine
• $40K (student tuition and stipendfor a year)
• Development of the method forpredicting signatures of the monitoredsystem, with one application area(semiconductor manufacturing dataset)
• 40K (student tuition and stipend fora year)
• Additional $20K for anotherapplication dataset (studentsupport and tuition for 6 months)
• Total: $80-100K (could be split into 2
years)