ABSTRACT The Korea Multi-purpose Accelerator Complex linear accelerator encompasses a Drift Tube LINAC (DTL) structure for high-frequency beam bunches accelerated by a radio frequency quadrupole. The existing anomaly detection system triggers shutdown when anomalies are detected. This study aims to detect anomalies to prevent unnecessary shutdowns through pre-adjustments. Drift-tube quadrupoles (DTQs) focus the beam within the DTL tank, but prolonged usage can degrade performance. To ensure stable beam focusing, the operational status of DTQ power supplies must be monitored. The EPICS input/output controller monitors these operational parameters and triggers alarms when thresholds are exceeded. However, this approach fails to consider changes, such as equipment aging. A machine learning-driven method can assimilate historical data related to the magnet power supply, enabling deriving adaptive thresholds, facilitating anomaly detection. A long short-term memory (LSTM) autoencoder model is employed to detect anomalies in DTQ magnets. Exploratory DTQ magnet operation data analysis was conducted to identify potential failures. The LSTM autoencoder offered a more adaptive and proactive detection model than conventional threshold-based methods. The results show that voltage fault events can be detected range from 30.7 to 286.6 minutes before they may cause a breakdown, enhancing device maintenance and performance through proactive prognostication of impending failures.
Read full abstract