Using Machine Learning Tools to Predict Accelerator Failure
M. Reščič, R. Seviour
University of Huddersfield, Huddersfield, United Kingdom
W. Blokland
ORNL, Oak Ridge, Tennessee, USA
Modern particle accelerator facilities are continuously introducing improved beam diagnostics, data acquisition, storage and analysis capabilities. Although the increased volume of data this generates makes manual analyse and understand the acquired data very difficult. In this paper we propose the use of machine learning to better understanding of beam failures. The proposed methods allow for both precognitive failure prediction and failure classification, determined using existing beam diagnostics infrastructure. Where we present the concept of tuning classifier parameters and pulse properties to refine datasets. As a demonstrator we apply our machine learning algorithm to analysis the vast data generated by the Oakridge Spallation Neutron Source (SNS) Differential Beam Current Monitoring (DBCM) diagnostics system. We show that analysis of the SNS DBCM data using machine learning, particle accelerator failure can be identified prior to the actual machine failure with 92% accuracy. Importantly, our research shows that emergent behavior regarding machine failure is encoded in the beam pulses prior to failure actually occurring.
Right click on video for Picture-in-Picture mode or Full screen display.