Particle accelerators are very complex facilities that are operated using thousands of devices generating vast amounts of data. Finding hidden patterns and correlations among this sea of information can be extremely challenging for human experts, and therefore machine learning techniques can be applied, gaining new understanding of the data and improving machine operation in many ways. Exploiting these techniques at the PSI accelerator facilities has been the target of the "Particle Accelerators and Machine Learning" (PACMAN) project, a collaboration with the SDSC (Swiss Data Science Centre). A safe Bayesian optimization routine has been applied to HIPA, which allows automatic safe and fast optimization of the beam losses. An interlock forecasting model that predicts beam interruptions by monitoring hundreds of diagnostic signals has been developed. In order to automatically monitor signals from beam diagnostics devices, and notify human experts if they behave in an anomalous way, a streaming anomaly detection system has been designed. Finally, regression models to replace failing diagnostic devices have also been developed and tested at HIPA.
For details, contact Jochem Snuverink