Anomaly Detection for Process Industry using Class-Dependent Temperatures Loss
Zhiqiang Li, Zhao Yang, Dong Mei Tang, Shaohu Peng, Jian Li · 2023
Anomaly detection is essential to the safety of control systems in the process industry. Thus, how to effectively detect anomalies in the presence of operational data has become a crucial issue. Currently, lots of studies utilize machine learning algorithms to detect anomalies in process industrial control systems. But most of these studies ignore the imbalance problem caused by the insufficient number of anomaly samples in the process industry. Therefore, this paper proposes an anomaly detection framework that introduces the Class-Dependent Temperatures (CDT) loss to construct the anomaly detection model. In the CDT loss, a weighting factor is adopted to alleviate data imbalance during the training phase. Our proposed model is evaluated on datasets collected from the Secure Water Treatment (SWaT) and the Water Distribution (WADI). Experimental results show that the method can effectively detect anomalies.