Anomaly Detection in Control Systems With Interval Dissimilarity

When performing anomaly detection on data generated by control systems, one is confronted with uniquely structured time series containing both set points as well as their corresponding process values. We devise a method which exploits the presence of these two distinct signals by converting time series to time interval series and subsequently applying a distance metric between pairs of time interval series to determine their anomaly status. Contrary to common time series distance metrics, the time interval series distance metric employed here scales with the number of intervals rather than with the number of samples in the time series. The distances produced by our approach respond appropriately to the occurrence of anomalous subsequences within the time series and are robust to minor variations between time series. The proposed method’s overall profile in terms of computational efficiency and quality of the produced distances is competitive in comparison with other methods in our experiments.