An automatic signal classification system for ultrasonic weld inspection signals

Neural network based automated signal classification (ASC) systems are bein g increasingly used to classify data obtained from nondestructive testing o f samples. This paper describes an ASC system using the Fuzzy ARTMAP networ k. Three important issues relevant to ASC systems, namely, incremental lear ning, confidence or reliability measures, and performance improvement, are studied. A fuzzy logic based algorithm is used to estimate the reliability of the network decision. The reliability of the classification decision is then incorporated into a feedback algorithm to improve the performance of t he network. Results on ultrasonic signals obtained from inspection of pipin g welds in nuclear power plants are presented.