Time-series Anomaly Detection Applied to Log-based Diagnostic System Using Unsupervised Machine Learning Approach

Annually, the Large Hadron Collider (LHC) demands a huge amount of computing resources to deal with petabytes of produced data. In the next years, a scheduled LHC upgrade will increase at least 10 times the computational workload on the Worldwide LHC Computing Grid (WLCG). As a consequence, an upgrade in the computing infrastructure that supports the physics experiments is also required. All WLCG computing centers are focused on the development of hardware and software solutions as machine learning log-based predictive maintenance systems. This work presents an original general-purpose diagnosis system to identify critical activity periods of services solving a binary anomaly detection problem using an unsupervised support vector machine approach.