Automating Safety Critical Ultrasonic Data Analysis with a Variational Auto-Encoder

Non-destructive evaluation based on the analysis of normal beam single probe pulse echo ultrasonic data is often a core component of a fitness for service argument supporting the continued use of manufactured components in safety critical contexts. Typical applications include testing of pipe walls, weldments, and concrete structure material condition. We model the analysis process with probabilistic statements and implement them by leveraging the empirically observed capacity of variational auto-encoders to learn and characterize complex distributions. The variational auto-encoder based characterization of ultrasonic datasets allows a partition that effectively removes outlier bias from consideration when applying deterministic methods for parameter estimation. This affords a bias free characterization of nominal response, the direct identification of deviations from it, and obtains real value estimates for the time of flight of features in the dataset. Our results are explainable, do not depend on out of distribution generalization, do not require data-preprocessing, and do not require massive labelled class balanced datasets. As such our method provides an ideal approach to support the automation of normal beam single probe ultrasonic data analysis in safety critical applications.