Application of convolutional and feedforward neural networks for fault detection in particle accelerator power systems

High voltage converter modulators (HVCM) provide power to the accelerating cavities of the Spallation Neutron Source (SNS) facility. HVCMs experience catastrophic failures, which increase the downtime of the SNS and reduce beam time. The faults may occur due to different reasons including failures of the resonant capacitor, core saturation due to the magnetic flux, insulated-gate bipolar transistor (IGBT) failures, and others. We recently have setup a HVCM test stand to develop and test machine learning models for anomaly detection and fault prognostics. In this work, we propose binary classifiers and autoencoder architectures based on convolutional (CNN) and feedforward neural networks (FNN) to facilitate distinguishing normal from faulty waveforms coming from the HVCM during operation. The results indicate that the CNN binary classifier is the best model among the four showing very stable performance in the training and testing sets with impressive precision and recall metrics, reaching up to 99% with a very small uncertainty. The FNN classifier shows the least performance with a large uncertainty in its metrics. The performances of the two autoencoders based on CNN and FNN were in between, showing very good performance nonetheless.