PNNL

Abstract

Ultrasonic testing (UT) is widely used in the nuclear industry to inspect thick steel components (i.e. pipes, welds, etc.) for the presence of potentially critical defects. The analysis of UT images requires highly trained experts and is both time consuming and costly. The use of machine learning (ML) applied to the analysis of UT data has shown its potential to deliver performances levels comparable to the those achieved by skilled operators, mitigating the errors without errors associated with so-called human factors [1][2]. In this paper, we aim to study how to optimize a ML training scheme to obtain a classification model that is sufficiently robust to generalize to different samples, given an inspection procedure (e.g. mono-element probe, TRL probe, PAUT). To achieve this, we propose to consider state-of-the-art classification algorithms (e.g., multilayer perceptron, support vector machine, extreme gradient boosting, deep neural networks, etc.) trained on small set of data comprising approximately 500 labelled experimental samples, coming, for each training stage, from one of the four weld specimens considered in this work. The testing procedure follows a cross-dataset approach, based on signals acquired from the remaining three specimens. The proposed analysis is based on the extraction of compact yet meaningful features from the C-scan echodynamic extractions, and on the automatic selection of ML models based on cross-validation. Comparisons of the analysis results against alternative data and features (for instance, B-scans and associated features) are also presented. The developed methodology allows us to establish a sufficiently robust pre-processing and training pipeline, which ensures good generalization performance in the cross-dataset testing scenario.