Quality assessment and anomaly detection for electronic storage ring injection based on machine learning

BackgroundDuring electron storage ring injection, beam parameters change drastically. Bunch-by-bunch three-dimensional position measurements can capture this transient process in real time to obtain multiple beam dynamics parameters. However, how to effectively utilize these measurements to evaluate and optimize accelerator operation remains a critical challenge for synchrotron radiation facilities.PurposeThis study aims to develop a machine learning-based method for injection quality assessment and anomaly detection in electron storage rings.MethodsFirstly, injection transient data accumulated over multiple years at the Shanghai Synchrotron Radiation Facility (SSRF) from September 2021 to June 2024 were selected as the sample database, including longitudinal oscillation amplitude, synchrotron damping time, bunch charge, and other dynamic parameters of refilled bunches. Then, the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm was applied to perform cluster analysis on the longitudinal oscillation amplitude and synchrotron damping time, identifying the temporal evolution patterns of facility operation states and labeling anomalous data samples that deviated from the main clusters. Finally, a k-Nearest Neighbor (k-NN) classification model was trained using the labeled dataset to predict anomalous injection events, with the optimal k value determined through comparative analysis.ResultsThe clustering results reveal the evolution patterns of SSRF operating states over time, with Cluster 2 (containing data from 2023‒2024) demonstrating significantly improved matching between the storage ring and injector compared to earlier periods. The average longitudinal oscillation amplitude decreases from 101.21 ps (before September 2023) to 64.37 ps (after September 2023), while the synchrotron damping time reduces from 3.01 ms to 1.70 ms. The optimized k-NN model with k=4 achieves a precision of 90%, recall of 66%, F1-score of 0.78, and PR-AUC of 0.84 in anomaly detection, outperforming other machine learning algorithms including Support Vector Machine, Decision Tree, Random Forest, Logistic Regression, and Naive Bayes.ConclusionsThe method proposed in this study combining unsupervised learning and supervised learning effectively detects anomalous injection events with 90% precision, enabling real-time quality assessment of storage ring injection performance. The longitudinal oscillation amplitude and synchrotron damping time serve as reliable indicators of injection quality, while the machine learning approach provides early warning capabilities for facility operators to prevent performance degradation before observable operational failures occur.