FAULT PROGNOSIS USING DYNAMIC WAVELET NEURAL NETWORKS P. Wang G. Vachtsevanos Georgia Institute of Technology Atlanta, GA AAAI 1999 Spring Symposium March 22-24, 1999
The ONR CBM Program Objective: Design, build and test an integrated hardware/software system for machinery diagnostics/prognostics The Testbed: A shipboard industrial chiller The Team: Honeywell Technology Center Predict DLI Georgia Tech NRL
The D/P/CBM Architecture System Time to Perform Maintenance FMEA Diagnostics Prognostics CBM Alarms
Diagnostics Determine accurately and without false alarms impending or incipient failure conditions
A Two-Prong Approach High-frequency failure modes (vibrations, etc.): The Wavelet Neural Net Approach Low-frequency events (Temperature, Pressure, etc.): The Fuzzy Logic Approach The Diagnostic Module
Failure Templates Fuzzify FeaturesInference Engine Fuzzy Rule Base (1) If symptom A is high & symptom B is low then failure mode is F1 (2)... (Defuzzify) Failure Mode Preprocessing and Feature Extraction Sensor DataFeatures
Wavelet Neural Network (WNN)
Wavelet Neural Network Competition Wavelet Neural Network Based Fault Classification Signal Feature Extraction Result + - Actual Fault Signature
Collect DataPreprocess data Extract Features & Prepare Training Data P 1 - peak of original signal P 2 - peak of its spectrum T - Fault Types C - Binary code of T P 1 P 2 C T Normal Vibrometer bias Strong Noise Bearing Crack.... P1P1 P2P2 C1C1 C2C2 Train WNN Input Data Preprocess Data Extract Features & Form Feature Vector WNN Result Off-line Learning On-line Implementation
Prognostics Objective –Determine time window over which maintenance must be performed without compromising the system’s operational integrity
The Prognosticator!
A Prognostic System
On Virtual Sensors Many failure modes are difficult or impossible to monitor Question: How do we build a “fault meter”? Answer: Virtual Sensor The Notion: Use available sensor data to map known measurements to a “fault measurement” Potential Problem Areas: How do we train the neural net? Laboratory or controlled experiments required
An Example (Ford Motor Co.) Engine Combustion Failures -- Misfire Detection Solution: Misfires may be discerned by detecting “acceleration deficits” Dynamic (Recurrent) Neural Net Crankshaft Acceleration Engine Speed Engine Load Misfire “Meter”...
Virtual Sensor
Process Demonstrator
An Experimental Setup
Width Depth
ON THE ISSUE OF UNCERTAINTY SOURCES OF UNCERTAINTY UNCERTAINTY REPRESENTATION UNCERTAINTY MANAGEMENT ONE-STEP VS. K-STEP PREDICTION CRITICAL CONCERNS ARISING IN FAILURE PROGNOSIS
Uncertainty boundaries in a prognostic task
Prediction of a distribution curve
Prognosticator using a fuzzy virtual sensor (Confidence) Interval DWNN
PREDICTION OF THE EVOLUTION OF A FAILURE AND ESTIMATION OF THE CONFIDENCE LIMIT UNCERTAINTY MANAGEMENT: SHRINK CONFIDENCE LIMITS
PREDICTION OF EVOLUTION OF FAILURE
Prediction models for sequence prediction (a) series-parallel model (b) parallel model
CONFIDENCE LIMITS USING RBFN OR WNN BASED PROGNOSTICATORS IF THE 95% CERTAINTY LIMIT FOR THE EXPECTED RESIDUAL VALUE ASSOCIATED WITH THE OUTPUT OF UNIT h IS: where t represents the student t-distribution and s is the local estimate of variance then: (Leonard, et al)
A POSSIBLE APPROACH TO SEQUENCE PREDICTION AND CALCULATION OF CONFIDENCE LIMITS
Current Research Thrusts Prognostics Uncertainty Management - Source of Uncertainty - Representation of Uncertainty - Uncertainty Management Learning/Adaptation Sensor Fusion
Issues and Concerns Large-grain uncertainty Lack of reliable failure models / failure growth prediction System-specific and operational-dependent conditions Incomplete and inadequate data sources Instrumentation and processing requirements Ability to deal effectively with multiple sensors / sensor fusion techniques Reliable, cost-effective, and efficient hardware platforms that can facilitate a parallel processing and multi-tasking environment
Conditioned-Based Maintenance Objective –Determine the “optimum” time to perform maintenance Problem Definition –A scheduling problem - schedule maintenance timing to meet specified objective criteria under certain constraints
Condition-Based Maintenance Major Objective –Extend system life cycle as much as possible without endangering its integrity Enabling Technologies –Various Optimization Tools –Genetic Algorithms –Evolutionary Computing