Source localization on large-scale canisters for used nuclear fuel storage using optimal number of acoustic emission sensors

Abstract

The dry cask storage system (DCSS) canisters have been used for the storage of high-level nuclear for decades. The inspections are needed to ensure that structural integrity is maintained. One mechanism of degradation on DCSS canisters that is of interest is stress corrosion cracking (SCC). Acoustic emission (AE) is a nondestructive technique that can be employed as an inspection approach since it can offer real-time degradation detection. This paper presents the approaches that can localize SCC sources by minimal acoustic emission (AE) sensor. To achieve this goal, three machine learning techniques (artificial neural network, random forest, stacked autoencoders) were adopted to improve the conventional source localization approach. In this paper, source localization is treated as a classification problem. The testing specimen was divided into multiple zones and located the AE signals to their corresponding zones. The AE signals were processed to create two datasets: a dataset consisting of AE parametric features and a dataset consisting of AE waveforms. Source localization approaches using artificial neural networks, random forest, and stacked autoencoders were trained and tested based on the datasets. The results show all three machine learning techniques can learn to map AE signals to their sources. Among them, stacked autoencoders have the best performance (97.8% accuracy of stacked autoencoders versus 91.5% of random forest, and 80.0% of ANN), demonstrating that it could be a potential approach to localize SCC events on DCSS canister.