Journal:Engineering Applications of Artificial Intelligence

Impact Factor:9.0

Summary:Anomaly detection in Industrial Control Systems (ICS) is vital for safety but challenged by complex data characteristics (multi-scale dynamics, hierarchical structures, time-varying dependencies, label scarcity) that impede existing methods. This paper introduces the Multi-Scale Causality-Inspired Convolutional Attention Network (MC-GCAN), a novel Interpretable framework for high-precision, interpretable ICS anomaly detection. MC-GCAN integrates hierarchical multi-scale feature learning for diverse temporal patterns and dependency-aware dynamic graph learning for adaptive sensor dependencies. It employs a refined dual-layer adaptive thresholding strategy for node/system-level anomaly determination, enhanced by contrastive learning for subtle anomalies. Extensive experiments on two benchmark datasets indicate that MC-GCAN achieves competitive performance compared to several state-of-the-art methods, demonstrating effective results and good generalization ability. Crucially, the framework demonstrates significant interpretability; validation confirms its ability to provide insights from system down to individual sensor levels. This node-level anomaly attribution holds substantial practical engineering value for root cause analysis. MC-GCAN offers a robust, effective, and interpretable solution for enhancing ICS security and reliability.

Key Words:Industrial control systems; Anomaly detection; Graph Neural Network; Multivariate time series; Adaptive thresholding

Indexed by:SCI JCR Q1

Note:https://www.sciencedirect.com/science/article/abs/pii/S0952197626015782

Discipline:Engineering

Document Type:JCR 一区

First-Level Discipline:Computer Science and Technology

Translation or Not:No