Anomaly Detection for the CERN Large Hadron Collider injection magnets

The Large Hadron Collider is the world’s largest single machine and the most powerful particle accelerator ever built. Inside it, high-energy particle beams are made to collide at speeds near the speed of light. The results of the collisions are then analysed in the hope that some of the fundamental open questions in physics can be answered. Particle beams are injected into the LHC and guided throughout it by thousands of electromagnets. New state of the art equipment for the LHC was developed by CERN in collaboration with thousands of scientists and engineers all over the world. The result of this is that, due to yet undiscovered reasons, equipment occasionally behaves in unexpected ways. This can cause experiments to fail, because of which the particle beams must then be ejected from the LHC so that the whole process of creating high-energy particle beams can be restarted. The behaviour that causes this is called anomalous behaviour. The goal of this thesis is to develop a method that can detect anomalies in the behaviour of injection kicker magnets, the magnets that inject particles into the LHC. A system that detects anomalies cannot be built manually, as there are over a hundred data signals related to the magnets and many complex relationships exist between the signals. Currently, anomalies are only detected at CERN after they have caused experiments to fail. A machine learning based application that can learn the complex relationships between signals and that can then detect anomalies automatically would thus greatly benefit CERN. Reaction times to anomalies could be improved and less time would need to be spent meticulously analysing signal data to find the causes of anomalies. Furthermore, it is possible for machine learning models to list the factors that contribute most to anomalous behaviour. Such information could be used by CERN to improve equipment. This thesis proposes an extensible anomaly detection application which supports the use of many different machine learning models. There is a limited amount of unspecific knowledge about when and where anomalies occurred in injection kicker magnet behaviour. A novel evaluation procedure is introduced that can use this unspecific knowledge to measure and compare anomaly detection performance fairly