Anomaly Detection Using Graph-Based Autoencoder with Graph Structure Learning Layer

The rapid technological advancements in manufacturing have necessitated sophisticated automation, control, and monitoring systems to enhance efficiency and productivity. These systems, characterized by multivariate time series data, often experience anomalies due to factors such as part malfunctions, cyber-attacks, and setting changes. Early detection of these anomalies is crucial for maintaining system safety, reducing repair costs, and ensuring high operational efficiency. However, the complexity of modern industrial systems frequently exceeds human monitoring capabilities, driving the need for advanced, data-driven anomaly detection (AD) models. This paper presents a novel approach to anomaly detection using a graph-based autoencoder with a graph structure learning layer. Our model dynamically learns and represents the relationships between variables in multivariate time series data, thereby enhancing the accuracy and reliability of anomaly detection. The key objectives include developing this model, improving AD performance, and validating real-world applicability through robust evaluation methods. Extensive experiments were conducted on real-world industrial datasets, including the SWaT and WADI datasets, which are benchmarks for evaluating AD models. The results demonstrated significant improvements in detection accuracy and reduced detection latency, validating the robustness and practicality of the proposed model. This research underscores the importance of a robust evaluation approach and highlights the critical role of detection latency in anomaly detection.