Anomaly detection for the injection kickermagnets of the CERN Large HadronCollider

Reliability, availability and maintainability are parameters that determine if a large-scale accelerator system can be operated in a sustainable and cost effective manner. Beam transfer equipment such as kicker systems are critical components with potential significant impact on the global performance of the entire machine complex. Identifying root causes of malfunctions is currently tedious, and will become infeasible in future systems due to increasing complexity. Machine learning can help toautomate this process. For this purpose a collaboration between CERN and KULeuven was founded. Several iterations of the study have yielded ananomaly detection pipeline which includes preprocessing, detection, postprocessing and evaluation. Merging data of different, asynchronous sources is one of the main challenges. Currently, Gaussian Mixture Models and Isolation Forests are used as unsupervised detectors, but any detector can easily be plugged into the system. During evaluation, we compare the detector predictions to manual e-logbook entries, which constitute a noisy ground truth. Lastly, we try to incorporate expert knowledge by means of semi-supervised clustering with COBRAS. A grid search allows for hyper-parameter optimization across the entire pipeline, yielding promising results