Computational Modeling for Automatic Superconducting Cavity Fault Prediction and Classification Using Time Series Signals

Processing multivariate time series signals collected from sensor networks is challenging because of complex temporal dependencies and non-stationarity. With the advent of artificial intelligence (AI) like machine learning and deep learning, it has become possible to process sensor-driven time series data more effectively than traditional statistical methods. This dissertation aims to develop machine learning and deep learning models to address machine fault diagnosis using multivariate time series signals collected from the Continuous Electron Beam Accelerator Facility (CEBAF) at Jefferson Lab. The first goal of the proposed work is to develop deep learning–based classification models and an unsupervised fault clustering approach using multivariate time series data. For the classification task, we designed a hybrid architecture that combines Long Short-Term Memory (LSTM) networks and Convolutional Neural Networks (CNN) to classify superconducting radio frequency (SRF) faults and cavity types. For the clustering task, we proposed a multivariate time series clustering approach based on Shapelet learning to identify fault patterns without requiring labeled data. The proposed model outperforms two other unsupervised clustering approaches, offering superior performance. The second goal of the proposed work aims to develop an SRF cavity fault prediction model using multivariate time series signals. In this work, we develop an uncertainty-aware hybrid deep learning model that combines LSTM and CNN architecture to predict SRF cavity fault using pre-fault signals. We incorporate multiple consecutive windows fault prediction and adjusting the fault confidence threshold strategies to enhance the fault prediction performance in real-time implementation. The model is evaluated using a highly imbalanced dataset from CEBAF, which reflects the real-world operating conditions. The results of the proposed approach may offer the potential to take preventive action before a failure occurs. The third goal of the proposed work aims to develop a domain adaptation network that performs fault predictions with changing data scenarios. We propose a Siamese network using a hybrid LSTM-CNN architecture to perform this task. The model performance is compared with another transfer learning model and a model with no transfer learning. The Siamese network-based domain adaptation approaches perform better than other approaches under changing machine dynamics.