Towards Decision Support and Automation for Safety Critical Ultrasonic Nondestructive Evaluation Data Analysis

Abstract

A set of machine learning techniques that provide decision support and automation to the analysis of data taken during ultrasonic non-destructive evaluation of Canada Deuterium Uranium reactor pressure tubes is proposed. Data analysis is carried out primarily to identify and characterizes the geometry of flaws or defects on the pressure tube inner diameter surface.

A baseline approach utilizing a variational auto-encoder ranks data by likelihood and performs analysis using Nominal Profiling (NPROF), a novel technique that characterize the very likely nominal component of the dataset and determines variance from it. While effective, the baseline method expresses limitations, including sensitivity to outliers, challenged explainability, and the absence of a strong fault diagnosis and error remediation mechanism.

To address these shortcomings, Diffusion Partition Consensus (DiffPaC), a novel method integrating Conditional Score-Based Diffusion with Savitzky-Golay Filters, is proposed. The approach includes a mechanism for outlier removal during training that reliably improves model performance. It also features strong explainability and, with a human in the loop, mechanisms for fault diagnosis and error correction. These features advance applicability in safety-critical contexts such as nuclear nondestructive evaluation.

Methods are integrated and scaled to provide: (a) a principled probabilistic performance model, (b) enhanced explainability through interpretable outputs, (c) fault diagnosis and error correction with a human-in-the-loop, (e) independence from dataset curation and out-of-distribution generalization (f) strong preliminary results that meet accuracy requirements on dimensional estimates as specified by the regulator in \cite{cog2008inspection}.

Though not directly comparable, the integrated set of methods makes many qualitative improvements upon prior work, which is largely based on discriminative methods or heuristics. And whose results rely on data annotation, pre-processing, parameter selection, and out of distribution generalization. In regard to these, the integrated set of fully learned data driven methods may be considered state of the art for applications in this niche context.

The probabilistic model, and corroborating results, imply a principled basis underlying model behaviors and provide a means to interface with regulatory bodies seeking some justification for usage of novel methods in safety critical contexts. The process is largely autonomous, but may include a human in the loop for fail-safe analysis. The integrated methods make a significant step forward in applying machine learning in this safety-critical context. And provide a state-of-the-art proof of concept, or minimum viable product, upon which a new and fully refactored process for utility owner operators may be developed.