Unsupervised Machine Learning for Fault detection in turn by turn Beam Position Monitors
Jakub Potaczala, IUT Orsay
Jeudi 16 juillet à 13h30, CEA-Saclay, Bat 130, pièce 102
Reporté
Turn-by-turn beam position monitors (TbT-BPMs) are essential for reconstructing the magnetic optics of high-luminosity colliders, yet faulty sensors and measurement noise routinely compromise data quality. As accelerators scale toward facilities like the Future Circular Collider (FCC), with thousands of BPMs distributed across a 91-km ring, manual inspection and hardware redundancy no longer suffice, motivating fully automated, unsupervised diagnostic pipelines. This seminar presents the improvements developed during an internship to a tool under development at CEA in collaboration with LISN at CentraleSupelec. The tool is used to study and compare complementary unsupervised machine learning approaches to TbT-BPM data from SuperKEKB's High and Low Energy Rings (HER/LER). A new pipeline including Isolation Forest, DBSCAN-based clustering coupled with two feature-engineering philosophies (a fully data-driven approach, Time2Feat, and a physics-informed fixed-feature approach) and a 1D convolutional autoencoder. Overall, these methods converge on a shared conclusion: unsupervised, physically interpretable pipelines can match or exceed manual fault curation, offering a scalable foundation for autonomous BPM diagnostics in next-generation accelerators.