Speaker
Description
The CERN Proton Synchrotron (PS) is equipped with several radiofrequency (RF) systems with heterogeneous hardware of different age. Despite high availability, some systems require frequent resets. Automatic fault recording supports post-failure root-cause analysis. However, interlocks alone may miss subtle beam or equipment degradation.
Machine learning-based anomaly detection has therefore been developed and deployed for the PS main accelerating systems. Autoencoders trained on cavity and beam signals provide a time-resolved anomaly score, complemented by rule-based checks. The models have already successfully highlighted degradations, enabling earlier interventions.
To handle the increased amount of data to review, a web-based interface presenting model outputs alongside cavity statistics was developed in 2026. Recognised signatures can be labelled, which allows adding explicit rule-based checks, progressively converting unsupervised detection into named failure modes. Together with the logged events accumulated in operation, these provide the labelled basis for supervised models, and a path from reactive fault analysis towards preventative maintenance of the PS RF systems.