Anomaly Detection Method for Complex Systems Based on a Domain-Adaptive Causal Decoupling Model
Chong Wang, Jie Liu
Abstract
Anomaly detection is vital for industrial system safety. The digitized industrial system has led to the widespread application of data-driven methods on anomaly detection modelling. Nevertheless, the time-variant operation conditions that change continuously the data distribution pose a significant risk of rendering models ineffective. To address this issue, a domain-adaptive causal decoupling model is designed in this article, considering the influence of operation conditions on system causality. And, an efficient anomaly detection method for complex systems is proposed, leveraging the domain-adaptive causal decoupling model. Specifically, a directed graph-based encoder is designed to capture the influence of the system operation modes and conditions on the causal structure and causal strength. Then, a causal decoupling model is proposed to learn domain-adaptive degradation status representations, which are essential for anomaly detection modeling. The effectiveness of the proposed method is verified via the real monitoring data of a high-speed train brake control system.