Evaluation of Machine Learning Tools for Inspection of Steam Generator Tube Structures using Pulsed Eddy Current
Inspection of multi-component systems, such as nuclear steam generator (SG) tube support structures, is complicated by multiple overlapping degradation modes. The simultaneous and precise measurement of more than two interdependent parameters is challenging when standard statistical regression analysis tools are used. Artificial neural networks (ANNs) have recently been applied to pulsed eddy current (PEC) data for inspection of Alloy 800 SG tube fretting, in the presence of tube off-set within a corroded ferromagnetic support structure. Signals were analyzed using modified principal component analysis (MPCA) followed by an ANN analysis, which simultaneously targeted four parameters associated with the support structure. These were hole diameter, tube off-centering in two mutually orthogonal directions and fret depth. In this work, the ANN analysis is compared with that performed by a Support Vector Machine (SVM) analysis of the same data. Comparable results are achieved for some parameters with both machine learning analysis tools. However, parameters with changing signal variance, such as those associated with support structure diameter, are not as easily compensated for using standard SVM analysis. Both techniques also rely on the availability of a representative training data set that may be difficult to come by for general inspection conditions.