INTEGRATION OF NEURAL NETWORKS AND EXPERT SYSTEMS IN MACHINE DIAGNOSIS

EXPERT SYSTEM TECHNOLOGY HAS BEEN PROVEN EFFECTIVE IN REACTIVE DIAGNOSIS, THAT IS, DIAGNOSING THE CAUSES OF MACHINE FAILURES AFTER THEY OCCUR. HOWEVER, EXPERT SYSTEMS HAVE BEEN LESS EFFECTIVE IN REPLICATING HUMAN EXPERTISE IN PROACTIVE DIAGNOSIS, THE PREDICTION OF IMPENDING FAILURES. ONE REASON FOR THIS LACK OF SUCCESS IS THAT PREDICATIVE DIAGNOSIS OFTEN DEPENDS ON FINDING FINDING PATTERNS IN VIBRATIONAL OR ACOUSTIC SIGNALS ("THE LIFTERS SOUND STICKY")RATHER THAN ON SYMBOLIC REASONING. CURRENT COMMERCIAL COMPUTER-BASED DIAGNOSIS SYSTEMS ARE NOT WELL SUITED TO DEALING WITH THESE SIGNAL PROCESSING TASKS, SINCE THESE EXPERT SYSTEMS USUALLY RELY ON EXPLICIT PROGRAMMING OR KNOWLEDGE ENTRY, AND EXPERTS RARELY KNOW HOW THEY PERFORM THESE RECOGNITION TASKS. HOWEVER, RECENT ADVANCES IN NEUROCOMPUTING HAVE RESULTED IN SYSTEMS WHICH LEARN TO RECOGNIZE PATTERNS IN SIGNALS WITHOUT BEING EXPLICITLY PROGRAMMED TO LOOK FOR THEM. THE WORK COVERED IN THIS PROPOSAL INVESTIGATES AN ARCHITECTURE THAT COMBINES THE SYMPTOM CLASSIFICATION ABILITY OF NEURAL NETWORKS WITH THE SYMBOLIC REASONING OF A DIAGNOSTIC EXPERT SYSTEM.