A Machine-Learning Based Methodology for Performance Analysis in Particles Accelerator Facilities

CERNs Accelerators facilities require adequate tools to operate, maintain and guide the consolidation of the machines, to fulfil the physics program and deliver the required luminosity for the experiments. It is therefore relevant to identify clear and easy metrics to monitor objectively the activities and guide the management to implement strategies, to improve performances, optimize costs as well as highlight key areas needing prioritization. Model-based is usually adopted to provide an early identification of failures and to reveal among several subsystems hidden dependencies. However system models are complex and require to be constantly updated to be reactive to system changes. Moreover, the availability of economic, high-efficient and smart sensors entails a growing data that, for volume and heterogeneity, are often difficult to analyse and to incorporate within the modelling process. The paper focuses on a novel approach which combines model-based and big data analytics principles. This involves machine learning techniques exploited to extract descriptive and predictive models directly from data, to foresee and react in time to the failures and reducing the downtimes of the machines which operate the largest particles accelerators complex of the world.