32b. Autoencoder Based Quench and Anomaly Detection in Superconducting Magnets

Superconducting magnet technology is one of the cornerstones of high energy accelerators. When small portions of these magnets cease to be superconducting, due to beam loss for example, the whole magnet can become normal conducting. This process referred to as a quench can be catastrophic. The ability to predict the onset of a quench and classify dierent quench events is highly valuable to the safe and reliable operation of superconducting accelerators. GENERAL STATEMENT OF HOW THE PROBLEM IS BEING ADDRESSED Our proposal will use machine learning to classify dierent types of quench events, identify quench precursors, and predict the onset of a quench. We will explore a range of unsupervised learning techniques and evaluate their ability to be deployed as part of real time systems using eld programmable gate arrays. WHAT IS TO BE DONE IN PHASE I? During Phaser I we will work with an accelerator laboratory to collect data necessary for building a machine learning application. We will then design classiers to identify dierent types of quench events. Following this we will work with high-speed data to both identify quench precursors and forecast the onset of a quench. We will integrate uncertainty quantication in to our software as well, in order to provide a condence interval for the predictions. Finally we will evaluate our ability to deploy these algorithms on a eld programmable gate array using specialized tools and compilers. COMMERCIAL APPLICATIONS AND OTHER BENEFITS The development of machine learning tools for quench prediction will fundamentally improve the performance and reliability of particle accelerators, while greatly reducing costs, both for the US high energy physics community and for many scientic facilities around the world. The methods developed here could also be applied to superconducting radio frequency cavities and other complex accelerator systems that have a tendency to fail. Moreover, machine learning for failure prediction is a highly valuable tool that has applications outside of accelerators.