Shapelet-Based Sensor Fault Detection and Human-Centered Explanations in Industrial Control System

Abstract

With the development of information and communication technology, industrial control systems (ICSs) that operate in a closed environment are now operating in a smart environment, and external threats are increasing. To predict failure and respond to threats, anomaly detection and fault detection using artificial intelligence (AI) are being introduced, but the issue of the reliability of AI prediction is emerging. In the case of anomaly detection, the operator must check thousands of sensors. In addition, since AI predictions are not always accurate, there are practical operational constraints, so this paper proposes shapelet-based anomaly detection and automatic failure sensor descriptions. When an abnormal situation occurs, the operator can immediately know which sensors caused the problem and how different the sensors appear from the normal patterns. This paper verified it with the HIL-based Augmented ICS Security Dataset (HAI) and Secure Water Treatment (SWaT) Dataset, widely used in the ICS field. In the case of the HAI Dataset, it was confirmed that 95.12% of the failed sensors could be analyzed by extracting and inspecting only 4% of the total sensors. In the case of the SWaT Dataset, only 7% of the total sensors were extracted and inspected, confirming that 84% of the failed sensors could be analyzed. We expect that intuitive explanations and anomaly detection will enable more effective technology operations in industrial environments.