1 © 1998 HRL Laboratories, LLC. All Rights Reserved Evaluation of Bayesian Networks Used for...
-
date post
21-Dec-2015 -
Category
Documents
-
view
213 -
download
0
Transcript of 1 © 1998 HRL Laboratories, LLC. All Rights Reserved Evaluation of Bayesian Networks Used for...
1© 1998 HRL Laboratories, LLC. All Rights Reserved
Evaluation of Bayesian NetworksUsed for Diagnostics[1]
K. Wojtek Przytula: HRL Laboratories
Denver Dash: University of Pittsburgh
Don Thompson: Pepperdine University
2© 1998 HRL Laboratories, LLC. All Rights Reserved
DIAGNOSIS - MODEL BASED APPROACH
DOMAIN MODELAPPLICATION DOMAIN
DIAGNOSIS SUPPORT TOOL
QUERY
DECISION
TROUBLESHOOTER
AnnumciatorPanel
Fuel Press FPS
Air Box P TBS
Fuel Temp FTS
Air Temp ATS
Crankcase CCP
Oil Temp OTS
Coolant P 1
Coolant P 2
Coolant P 3
Oil P PresOPS
PowerWiring
HarnessEM 2000
Power Supply
Comm Interface
& Fault Data
EM 2000Wiring
Harness
FuelPump
PrimaryFilter
SensorWiring
Harness
Fuel Injectors9 - 16
LeftInjectorHarness
ECM Sender
Crankshaft Spd SRS
Crankshaft Pos TRS
ECM ReceiverRight
InjectorHarness
Fuel Injectors1 - 8
PerformanceSensors
Protective System
ABCD 74 VDC 24 VDC
SuctionStrainer
Thermo Mod AMOT Valve
FuelTank
30 psiBypass
Fuel Preheater
120 PSIRelief Valve
SecondaryFuel Filters
SensorComponent
Wiring Harness
Pipe
Wiring
FuelSystem
40 psi Relief Valve
Cold Plate
Fuel Pressure
To Fuel Tank
3© 1998 HRL Laboratories, LLC. All Rights Reserved
Bayesian Network Diagnostics
•Bayesian Networks as models for computerized diagnostic assistants
•Model evaluation has not been addressed
•Model quality determines diagnosis quality •Evaluation provides a basis for model performance estimation
4© 1998 HRL Laboratories, LLC. All Rights Reserved
GRAPHICAL MODEL FOR DIAGNOSIS (GRAPH AND PROBABILITY THEORY)
GRAPH (structure):
• Two Fault Nodes: F1, F2.
• Three Observation Nodes – e.g. symptoms and tests.
• Causal Links
PROBABILITIES (parameters)
• Prior Probabilities of Faults
• Conditional Probabilities of Observations given Faults
• The Model constitutes a joint probability distribution over the nodes.• It is obtained from data or knowledge or both.
6© 1998 HRL Laboratories, LLC. All Rights Reserved
Bayesian Network Evaluation]
Using Inference, Monte Carlo Simulation, & Visualization Techniques
–Step 1–Set Defective Component–Execute Forward Inference
–Step 2–Sample Observation States–Execute Reverse Inference
9© 1998 HRL Laboratories, LLC. All Rights Reserved
Identification of Critical Elements Responsible for Incorrect Diagnosis
–Components with weak observations that cannot be diagnosed convincingly
–Strongly coupled components that implicate each other, so they cannot be effectively separated in diagnosis
–Components whose failures are misinterpreted as failures of other components
Evaluation Conclusions
10© 1998 HRL Laboratories, LLC. All Rights Reserved
Fuel Level Battery Starter CableFuel PumpFuel Filter Induction Coil
Sample Graph for Car Diagnosis Bayesian Network Model[1]
11© 1998 HRL Laboratories, LLC. All Rights Reserved
2-D Matrix for Car Diagnosis Bayesian Network Model1]
Starter
Fuel Level
Battery
Cable
Fuel Pump
Fuel Filter
Induction Coil
Prior Probabilities
True Defect
IMPLICATED
FAULT
12© 1998 HRL Laboratories, LLC. All Rights Reserved
3-D Matrix for Car Diagnosis Bayesian Network Model 1]
Prior Probabilities
Starter
Fuel Level BatteryCable
True Defect
IMPLICATED
FAULTFuel Pump
Fuel FilterInduction Coil
Starter
Fuel Level
Battery
Cable
Fuel Pump
Fuel Filter
Induction Coil
13© 1998 HRL Laboratories, LLC. All Rights Reserved
3-D Matrix for Bayesian NetworkModel for the Large Network [1]
14© 1998 HRL Laboratories, LLC. All Rights Reserved
Results and Conclusions [1]
RESULTS
• METHOD AND ALGORITHMS FOR ANALYSIS OF BAYESIAN NETWORKS FOR DIAGNOSTICS
• SOFTWARE PACKAGE FOR COMPUTATION AND DISPLAY OF THE ANALYSIS RESULTS
CONCLUSIONS
• THE RESULTS CAN BE USED AS A GUIDE IN TESTING OF THE MODEL
• THE METHOD CAN BE USED IN DESIGN OF SYSTEMS FOR DIAGNOSIBILITY
• THE METHOD IS APPLICABLE NOT ONLY TO DIAGNOSTICS
BUT TO GENERAL CLASS OF DECISION SUPPORT PROBLEMS