Explainable Deep Learning for Fault Prognostics in Complex Systems: A Particle Accelerator Use-Case

Sophisticated infrastructures often exhibit misbehaviour and failures resulting from complex interactions of their constituent subsystems. Such infrastructures use alarms, event and fault information, which is recorded to help diagnose and repair failure conditions by operations experts. This data can be analysed using explainable artificial intelligence to attempt to reveal precursors and eventual root causes. The proposed method is first applied to synthetic data in order to prove functionality. With synthetic data the framework makes extremely precise predictions and root causes can be identified correctly. Subsequently, the method is applied to real data from a complex particle accelerator system. In the real data setting, deep learning models produce accurate predictive models from less than ten error examples when precursors are captured. The approach described herein is a potentially valuable tool for operations experts to identify precursors in complex infrastructures.