Diffusion Partition Consensus: Diffusion-Aided Time-of-Flight Estimates, Anomaly Detection, and Localization for Ultrasonic Nondestructive Evaluation Data

Abstract

Diffusion partition consensus is a novel generative AI-based technique for time-series anomaly detection and data imputation in the presence of outliers. To illustrate the method, an implementation with design choices tailored for well-structured time series typical of single probe ultrasonic nondestructive evaluation (NDE) datasets is proposed. The technique relies on cross-talk between a conditional score-based diffusion model, and two well-chosen Savitzky-Golay filters. Testing and evaluation are performed on a series of progressively information rich synthetic datasets, and on real-world ultrasonic NDE datasets taken from a Canada Deuterium Uranium nuclear reactor pressure tube and calibration fixture. The iterative technique is a blend of stochastic and deterministic methods that uses confidence and consensus of target parameter estimates to update several data classifying partitions over the dataset, which in turn allows a new set of estimates and confidence measures to be established. Data classification induces a progressive bias in the training datasets allowing a diffusion model to identify the prevalent distribution. Methods for fault diagnosis support the efficacious inclusion of a human in the loop making the technique suitable for use in safety-critical applications. The main advantages of the technique are that it is unsupervised—in that it does not require labeled datasets or significant data preprocessing, does not rely on out-of-distribution generalization, provides means for fault diagnosis without recourse to ground truth, converges with stability, and naturally includes a human in the loop. The quality of results, the checks and balances provided by the fault diagnosis mechanism, and the opportunity to include a human in the loop, support the case for usage in safety-critical contexts such as NDE at a nuclear power facility.