Eddy current flaw characterization in tubes by neural networks and finite element modeling

Abstract

For quantitative flaw characterization in steam generator tubes, inversion of eddy current testing (ECT) signals in an automated fashion is strongly desired. In this paper, we report our effort to develop a systematic approach for flaw characterization in tubes by the novel combination of neural networks and finite element modeling. Specifically, the finite element model that can predict ECT signals from axisymmetric flaws in tubes was developed, and its accuracy was verified experimentally. Using this model, an abundant synthetic database with 400 ECT signals generated from 200 axisymmetric machined grooves in four types has been constructed with two test frequencies per flaw. For the automated inversion of ECT signals, a total of 22 features have been extracted from each flaw. Then, a set of 10 features has been selected for flaw classification, while the other set of 10 features for flaw sizing. For the determination of the flaw type and the flaw size parameters, we have proposed an intelligent flaw characterization system that adopts two different paradigms of neural networks: probabilistic neural networks for flaw classification and back propagation neural networks for flaw sizing. The performance of this system has been investigated using the synthetic ECT signals in the database. The excellent performance presented here, even though it has been obtained from synthetic flaws representing machined grooves in tubes, demonstrates the high potential of this system to serve as a robust tool for practical flaw characterization in tubes.