An Anomaly Detection Approach for Realtime Identification Systems Based on Centroids

The present research describes a novel adaptive anomaly detection method to optimize the performance of nonlinear and time-varying systems. The proposal is based on combining the real-time identification algorithm, Recursive Least Squares, with a centroid-based methodology. For anomaly detection, the method compares the current system dynamics with the average (centroid) of the dynamics identified in previous states for a specific setpoint. If the dynamics difference exceeds a certain threshold, the system classifies it as an anomaly. Otherwise, the centroid is updated by introducing the newly identified data. Finally, the proposed method was tested on a real system, in this case, on the level control plant, obtaining a good performance in anomaly detection.

Automated summary

The present research proposes a novel adaptive anomaly detection method for optimizing the performance of nonlinear and time-varying systems. The method combines the real-time identification algorithm Recursive Least Squares with a centroid-based approach. For anomaly detection, the method compares the current system dynamics with the average dynamics identified in previous states for a specific setpoint. If the dynamics difference exceeds a certain threshold, it is classified as an anomaly. Otherwise, the centroid is updated with the newly identified data. The proposed method shows good performance in anomaly detection when tested on a real system, specifically the level control plant.