Development of fault identification pipeline for SPIRAL2 LLRF data

TL;DR

This paper addresses the challenge of interpreting faults in SPIRAL2 LLRF time-series data by applying a Time2Feat-based pipeline to extract informative features from multivariate signals. It implements a configurable clustering and visualization framework, enabling interpretable grouping of anomaly patterns and facilitating an eventual real-time state observer. A case study differentiates electronic quenches from false quench alarms using tsfresh features, dimensionality reduction via OpenTSNE, and DBSCAN clustering, supplemented by a KNN classifier with promising results on a small labeled set. Although not yet delivering robust universal classification, the approach provides a scalable path toward faster fault identification and reduced data storage through selective filtering, with clear directions for improving interpretability and real-time deployment. The work advances FAIR data practices for SPIRAL2 LLRF postmortem data and lays groundwork for more automated and explainable anomaly detection in accelerator RF systems.

Abstract

SPIRAL2 is a state-of-the-art superconducting linear accelerator for heavy ions. The radiofrequency operation of the linac can be disrupted by anomalies that affect its reliability. This work leverages fast, multivariate time series post-mortem data from the Low-Level Radio Frequency (LLRF) systems to differentiate anomaly groups. However, interpreting these anomalies traditionally relies on expert analysis, with certain behaviours remaining obscure even to experienced observers. By adopting the Time2Feat pipeline, this study explores the interpretability of anomalies through feature selection, paving the way for real-time state observers. Clustering dashboards are presented, allowing the use of multiple clustering algorithms easily configurable and tools to help for visualizing results. A case study on distinguishing electronic quenches and false quench alarms in postmortem data is highlighted. Thereby, a fast and reliable K-Nearest Neighbours (KNN) classifier is proposed.

Figures (8)

• Figure 1: SPIRAL2 cryomodules.
• Figure 2: Acquisition system overview.
• Figure 3: Summary of events by year.
• Figure 4: Evolution of alarms over time.
• Figure 5: Example of clustering results with upset plot.