Time Delay Neural Networks for Classification of Ultrasonic NDT Signals

The pressures of marketplace are increasingly resulting in a shift in emphasis from such concepts as Scheduled Replacement to notions that include Retirement for Cause (RFC). The shift has resulted in enormous demands placed on the performance levels of NDE algorithms and systems. The need for detection of flaws in the incipient stages of growth and the regularity with which inspections need to be carried have contributed to interest in systems that offer high levels of accuracy and throughput. Pattern recognition algorithms play a key role in NDE systems and improvements in classification schemes are an area of active research. This paper presents an approach for addressing classification of ultrasonic signals obtained during the inspection of nuclear power plant tubing. The signals are analyzed for detection of intergranular stress corrosion cracking (IGSCC). However the similarity of the IGSCC signals to reflectors from other benign sources such as rootwelds and counterbores renders the identification of crack signals difficult. In addition variations in the location of cracks are manifested as temporal shifts or delays in the signal. The classification scheme is therefore required to be insensitive to temporal shifts.