Speaker
Description
Real-time bunch-by-bunch diagnostics are essential for stable synchrotron operation under top-off conditions. Conventional monitoring based on individual observables struggles to characterize complex beam dynamics from multi-parameter coupling. To address this, an unsupervised deep-learning framework was developed for automatic longitudinal bunch-by-bunch signal diagnostics and machine parameter monitoring at the Hefei Light Source II (HLS-II). The framework utilizes high-precision steady-state bunch phase measurements through two complementary data-driven branches. An autoencoder (AE) branch with dimensionality reduction and ensemble clustering characterizes the global structure of machine operating states. In parallel, an LSTM-VAE branch captures temporal correlations in longitudinal oscillation signals, enabling localized anomaly detection via multi-scale reconstruction error analysis and an Isolation Forest. Applied to HLS-II operational data, the framework successfully identified a significant longitudinal anomaly on September 18, 2025. Correlation with historical machine parameters revealed that the event was driven by large phase fluctuations in the higher-harmonic cavity, which modified the longitudinal potential well and altered the synchrotron frequency. The results demonstrate that the proposed method effectively identifies latent parameter anomalies and their underlying physical origins without labeled data, providing a practical tool for online beam diagnostics.
| I have read and accept the Privacy Policy Statement | Yes |
|---|---|
| Supervisor's name | Yongbin Leng |
| Supervisor's email | lengyb@ustc.edu.cn |
| Paper submission software | LATeX |