Extension of Signal Monitoring Applications with Machine Learning

The Large Hadron Colider (LHC) is the world’s largest particle accelerator. It is 27-km long and contains a wide range of superconducting circuits for controlling the shape and trajectory of particles. During operation, the nominal designed current (for 7 TeV) in the main bending dipole circuit is 11 850 A, which is equivalent to the current of about 120 single-family households. In order to prevent failures during operation, there are several protection systems installed. Furthermore, each of the magnets is checked during the Hardware Commissioning (HWC) powering test, which take place prior to each operation following an extended technical stop. Especially, because of the high complexity of the LHC and the requirement of high reliability during operation, those safety measures have a huge responsibility. Many protection systems have taken care of this responsibility in the past, which led to several years of successful operation. The data gathered during these years, allows the characterisation of the protection systems and the usage of the obtained values as reference for the monitoring during operation. The "LHC Signal Monitoring Project" has been founded to unite existing analysis tools. This thesis shows how the logged signals of the different databases can be used in order to implement new and extend existing monitoring applications into the development environment of the project. Several LHC component features are calculated and their significance is discussed. Since the LHC consists of several copies of similar circuits, the distribution of those parameters is studied and compared over both time and circuit. In particular, the implementation of two existing LHC analysis modules from the past are presented in this thesis. The busbar resistance analysis and the quench heater analysis. For both methods, this thesis provides a generic analysis which can be applied to any signal of the LHC systems. It covers the data analysis steps of acquisition, exploration, modelling, and monitoring. During modelling a unique approach is introduced, which uses supervised machine learning to extend existing signal monitoring applications