Deep learning for anomaly detection in high-energy beam dump data from the Large Hadron Collider

The Large Hadron Collider (LHC) is the world’s largest and most powerful particle accelerator, in which beams of particles are accelerated to near the speed of light. Partial losses of high-energy beams can damage crucial LHC equipment, requiring tedious and costly reparations. Beam loss monitoring is performed in order to ensure the safe continuous operation of the LHC required for studying rare phenomena in high-energy physics. In complement to automatic system checks, a manual inspection of beam losses by a machine protection expert is performed after every high-energy beam dump in the LHC, to allow the early identification of issues. In this master’s thesis, we explore two different data-driven approaches to automatizing the analysis of beam losses during high-energy beam dumps in the LHC, with the aim to facilitate the analysis of future beam dumps and contribute to the understanding of beam loss mechanisms. In the two approaches, methods for anomaly detection are applied to historical beam dump data from several years of successful operation. For the first approach, we develop and test a physics-model–driven method to classifying beam dumps, where beam losses are modeled using regression analysis and expert knowledge about physical phenomena in the LHC. For the second approach, we develop and test a classification method using deep learning, for the purpose of efficiently extracting relevant information from large beam dump datasets. Building on this method, we also propose a method using variational autoencoders, for which several different anomaly measures can be defined. Classification results from the physics-model–driven method were validated by machine protection experts, and show that the method works very well for detecting anomalous beam losses. We propose several possible improvements to this promising method, that could lead to its full implementation into the existing beam loss diagnostics systems. Results from the classification method using deep learning show that it can detect known types of beam loss anomalies, and proves the capability of the neural network architecture. This provides a foundation for future studies on the subject, for which we identify several possibilities.