Early Fault Detection in Particle Accelerator Power Electronics Using Ensemble Learning

Early fault detection and prognosis are crucial to ensure efficient safe operations of complex engineering systems such as the Spallation Neutron Source (SNS) its power electronics (high voltage converter modulators). Following an advanced experimental facility setup that mimics SNS operating conditions, authors successfully conducted 21 early experiments, where precursors introduced in system a degree enough cause degradation waveform signals, but not reach real fault. Nine different machine learning techniques based on ensemble trees, convolutional neural networks, support vector machines, hierarchical voting ensembles proposed detect precursors. Although all 9 models have shown perfect identical performance during training testing phase, most has decreased next test phase once they got exposed realworld data from experiments. The ensemble, which features multiple layers diverse models, maintains distinguished with 95% success rate (20/21 tests), followed by adaboost extremely randomized trees 52% 48% rates, respectively. were worst only 24% (5/21 tests). study concluded successful implementation or particle accelerator would require major upgrade controller acquisition facilitate streaming handling big for models. In addition, this shows best performing concept reduce bias hyperparameter sensitivity individual