Machine-learning-based pressure-anomaly detection system for SuperKEKB accelerator

Abstract

This study developed a pressure-anomaly detection system utilizing machine learning for the vacuum system of the SuperKEKB accelerator. The system identified abnormal pressure behaviors among approximately 600 vacuum gauges before triggering the conventional alarm system, facilitating the early implementation of countermeasures and minimizing potential vacuum issues. By comparing the recent pressure behaviors of each vacuum gauge with the previous behaviors, the program detected anomalies using the decision boundary of a feed-forward neural network previously trained on actual abnormal behaviors. Realistic regression models for pressure data curves enabled a reasonable prediction of the causes of anomalies. The program, implemented in python, has been operational since April 2024. Although based on a rudimentary machine-learning concept, the developed anomaly detection system is beneficial for ensuring the stable operation of large-scale machines, including accelerators, and is helpful in designing systems for fault detection.