The inspection of butt-welded stainless-steel pipe joints in nuclear power plants is routinely performed using ultrasonic non-destructive evaluation methods. Field experience, based on conventional ultrasonic signal amplitude criteria, indicates that a large number of indications are recorded. Most of these are not due to cracks, but are inherent in the geometry of the specimen. Discrimination between crack and geometrical/weld (malignant vs. benign) indications is principally based on operator experience, variations in signal amplitude, and the location of the reflector. Field experience and round-robin tests show that indication discrimination is a very time-consuming process. Besides, significant variations in performance exist due mainly to operator experience, fatigue, concentration, and conventional signal amplitude evaluation criteria.

In response to this problem, this paper describes an artificial intelligence methodology and results for classification of intergranular stress-corrosion cracking (IGSCC) from geometrical/weld reflectors in austenitic stainless-steel pipes. This algorithm was developed using the protocol method of artificial intelligence heuristic programming and, as such, can provide answers comparable to those supplied by well-trained technicians during the flaw discrimination process, i.e., discrimination between crack and geometry/weld ultrasonic signals. Preliminary results show that this approach yields a better than 90-percent index of performance.