| Abstract [eng] |
Industrial control systems must be running as stable as possible. However, in such processes, due to certain internal or external factors, anomalies may occur. Such anomalies are difficult to detect because of the complexity of the processes and large volume of process data. Therefore, various machine learning methods are used for their detection. In this work, five different machine learning models are implemented and trained to detect anomalies in water treatment system data. The study uses the “SWaT” and “WADI” datasets, collected from real water treatment systems, which are preprocessed before being used for model training. Optimization methods are applied to select the most suitable model parameters. Model performance is evaluated using both standard metrics, such as the F1 score, and range-based metrics, which assess model effectiveness over time intervals. As established in the study, this is a more appropriate method for evaluating performance in industrial process data. After completing the necessary steps to evaluate the performance of the investigated machine learning models, conclusions about the selection of the most suitable model for anomaly detection in water treatment system data were made. |
