Fusing Design and Machine Learning for Anomaly Detection in Water Treatment Plants

Abstract

Accurate detection of process anomalies is crucial for maintaining reliable operations in critical infrastructures such as water treatment plants. Traditional methods for creating anomaly detection systems in these facilities typically focus on either design-based strategies, which encompass physical and engineering aspects, or on data-driven models that utilize machine learning to interpret complex data patterns. Challenges in creating these detectors arise from factors such as dynamic operating conditions, lack of design knowledge, and the complex interdependencies among heterogeneous components. This paper proposes a novel fusion detector that combines the strengths of both design-based and machine learning approaches for accurate detection of process anomalies. The proposed methodology was implemented in an operational secure water treatment (SWaT) testbed, and its performance evaluated during the Critical Infrastructure Security Showdown (CISS) 2022 event. A comparative analysis against four commercially available anomaly detection systems that participated in the CISS 2022 event revealed that our fusion detector successfully detected 19 out of 22 attacks, demonstrating high accuracy with a low rate of false positives.