Feasibility of using adaptive learning networks for eddy current signal analysis. Final report. [PWR

Abstract

Unambiguous discrimination and accurate sizing between simulated pits and cracks have been obtained via Adaptive Learning Network (ALN) flaw-classification and ALN flaw-size models for both single and multiple frequency eddy current data. In terms of sizing flaws, the error rates were 2.4 percent for pits and 3.6 percent for cracks. Eddy current signal responses were generated, recorded, and digitized from several simulated pits and cracks in sample nuclear reactor steam generator tubing. These responses were parameterized to measure the in-phase and quadrature signal and power components. It is concluded from this feasibility study, which considered 100, 200, 300 and 400 kHz signals, that the optimum inspection mode for pits is with a single frequency (400 kHz) eddy current carrier signal and the optimum inspection mode for cracks is with multiple frequencies (200 kHz and 400 kHz).