Operational data for fault prognosis in particle accelerators with machine learning

Abstract

This paper presents real operational data collected from the power systems of the Spallation Neutron Source facility, which provides the most intense neutron beam in the world. The authors have used a radio-frequency test facility (RFTF) and simulated system failures in the lab without causing a catastrophic system failure. Waveform signals have been collected from the RFTF normal operation as well as during fault induction efforts. The dataset provides a significant amount of normal and faulty signals for the training of statistical or machine learning models. Then, the authors performed 21 test experiments, where the faults are slowly induced into the RFTF system for the purpose of testing the models in fault prognosis to detect and prevent impending faults. The test experiments include interesting combinations of magnetic flux compensation and start pulse width adjustments, which cause gradual deterioration in the waveforms (e.g., system output voltage, system output current, insulated-gate bipolar transistor currents, magnetic fluxes), which mimic the fault scenarios. Accordingly, this dataset can be valuable for developing models to predict impending fault scenarios in power systems in general and in particle accelerators in specific. All experiments occurred in the Spallation Neutron Source facility of Oak Ridge National Laboratory in Oak Ridge, Tennessee of the United States in July 2022.

Keywords

• Fault prognosis
• Health management
• Machine learning
• Signal processing
• Spallation neutron source
• High voltage converter modulators