An Active Reference Reset Method Adapting Distribution Shift for Robust System Anomaly Detection
Seungwan Seo, Heejeong Choi, Pilsung Kang · 2022 13th International Conference on Information and Communication Technology Convergence (ICTC) · 2022
In order to ensure the operation reliability of process systems in various industries, faults and unplanned shutdowns must be detected early. Traditionally, they have been detected based on the domain knowledge of industrial engineers. However, data-based methods have been developed with recent advances in machine-learning algorithms. These methods can detect signs of shutdowns and set the alarms off. However, they can not account for data distributions that have undergone changes after shutdowns. To handle this challenge, we propose a robust system anomaly detection algorithm using the distribution-based active reference set reset method. The experimental results on real-world industry data show that the proposed method could detect abnormal signs early at various thresholds.