Using Machine Learning to Predict Failures in a Particle Accelerator
Timur Gulur, MacKenzye Leroy, Colin O'Brien, Ryan Pindale
Large-scale instruments are vital to the progression of scientific discovery. Instrument downtime often stalls research; by reducing downtime, experimenters can increase research productivity and attain higher returns on investment. Our team focused on instruments of high complexity, where electrical issues in various subcomponents have the potential to cause problems ranging from simple experimental failure to catastrophic system damage. We propose a novel approach for preemptive detection of electrical faults using a variety of machine learning methods on signal data from Oak Ridge Laboratory’s Spallation Neutron Source (SNS) particle accelerator. We compared four methods: a prototypical network that uses Symbolic Fourier Approximation for feature engineering and few shot learning for training, a Gaussian Process Classifier, an Approximated Bayesian Neural Network using Monte Carlo Dropout, and an LSTM Autoencoder. We evaluate these methods based on their ROC curves and provide a general commentary on the advantages and disadvantages of each method. Our results demonstrate capacity for identifying the imminence of certain failure states and provide avenues for future enhancement.