Using machine learning to automate seismoacoustic process monitoring of a nuclear pyroprocessing element chopper

Introduction: Advancements in novel sensors and artificial intelligence data processing have increased the prevalence of condition and process monitoring across various industries, including the nuclear sector. This study evaluates the feasibility of using machine learning (ML) techniques applied to seismoacoustic signals for monitoring pyroprocessing activities, using data collected from three sensors installed near equipment performing nuclear fuel chopping at Idaho National Laboratory. The study offers two main research contributions: (1) an assessment of the feasibility of training a classification model and testing it with acoustic data from the same sensor as was used for training (termed “in-domain performance”), and (2) an investigation into the feasibility of generalizing learned patterns, allowing a model trained on data from one or a combination of sensors to be tested on data from different sensors on the same type of equipment (termed “out-of-domain performance”). Methods: Two high-level approaches were considered: engineered features using the short-time Fourier transform (STFT) coupled with extreme gradient boosting (XGBoost), feedforward neural network (FNN), and Isolation Forest (iForest) classifiers; and learned features using convolutional neural networks (CNNs). Additionally, for the second contribution, an advanced deep learning framework called domain-adversarial convolutional neural network (DACNN) was used, having been designed for domain adaptation applications. Results: The results demonstrated highly successful in-domain performance, with the best-performing model XGBoost achieving a precision of 93%, a recall of 92%, and an F1 score of 94%. However, out-of-domain performance was lower, with precisions of 53%–69%, recalls of 53%–86%, and F1 scores of 56%–69% for the best performing models, thus highlighting the challenges of generalizing to new sensors. Discussion: These findings are significant for the nuclear pyroprocessing industry, as they illustrate the feasibility of using ML to automate seismoacoustic process monitoring of a nuclear pyroprocessing element chopper, affording potential benefits such as improved safety, enhanced material safeguards, and increased operator awareness.