Anomaly Detection for the European XFEL using a Nonlinear Parity Space Method

Abstract

The European XFEL is currently the largest linear accelerator using superconducting technology with 808 electromagnetic field cavities. The existing fault detection system focuses on the most severe known faults using heuristically tuned thresholds on selected physical parameters computed from measurements. In the large amount of cavity data, non-optimal and abnormal behavior is not always classified. The aim of this article is to describe an automatic anomaly detection system for the European XFEL to detect and store anomalies in the cavity behavior for further analysis. A parity space algorithm using a nonlinear model for the nominal behavior of each cavity is proposed. Due to its low computational complexity, this algorithm is suitable for an implementation on the existing data acquisition and control system based on FPGAs. The proposed fault detection scheme allows the classification of numerous data sets into faults, anomalies and nominal behavior. The evaluation on available data sets shows that a further distinction between anomalies and faults is possible, useful and might even lead to new insight into the formation of (possibly severe) faults during the operation of the system.