AbstractIn this thesis, the use of machine learning methods for predicting particle accelerator failures is presented. The methods predict failures ahead of time providing time to take action before the failures occur. The aim of the method is to raise the particle accelerator reliability as new mitigation techniques are enabled by this approach. A literature study on reliability of particle accelerators, their applications and associated issues was performed and the need for a novel approach established. An evaluation of non machine learning methods from applicable industries, Markov Models and Electric Network Frequency Criterion (ENF), was performed. Overview, feasibility studies and performance evaluations on most suitable machine learning methods are presented. Potential issues and limitations are presented and addressed. Methods from the machine learning domain are studied and evaluated. The feasibility of binary classifiers as the most suitable tools is researched and established. The most suitable classifiers are evaluated and validated for real world applications at Spallation Neutron Source (SNS) using standardized performance metrics from the machine learning domain. All measurements are performed and validated on two independent datasets acquired from SNS. Requirements for implementation are presented and discussed. Various optimizations for implementation in hardware are presented. A binary classifier, meeting the real-world application speed requirements, capable of predicting 58% accelerator failures where beam loss occurs while maintaining a 0% error rate is presented. The methods meet the requirements to be implemented and used as a mitigation technique for inhibiting pulse generation prior to accelerator failure. SNS is implementing the method in their real-time Field Programmable Gate Array (FPGA) hardware as a part of the Machine Protection System (MPS).
The implementation and evaluation of the tools for the research were written in python using the libraries scikit-learn and numpy. The code was shared with SNS where it is used to port the methods to FPGA and to independently verify the results.
|Date of Award||5 Aug 2022|
|Supervisor||Rebecca Seviour (Co-Supervisor)|