Capturing principal features in slow industrial processes for anomaly detection application

Abstract

Dynamics is the fundamental characteristic of running industrial processes, and online industrial anomaly detection requires satisfactory real-time property and precision. Uncertain noise and outliers are also common in industrial data, sharply decreasing the performance of data models. Moreover, the embedded dimension has been a consensus for high-dimensional processes because several factors drive the plant. Based on these rationales, we propose a new dynamic anomaly detection strategy named robust principal slow feature analysis (RPSFA). This method could preserve the local geometric structure and is robust to outliers. Moreover, the proposed method effectively realizes information-noise separation to improve detection performance. Two pairs of detection statistics are constructed to concurrently monitor the process’s steady state deviation, dynamic characteristics change, noise anomaly, and the breakdown of variable correlations. A numerical case and a simulated industrial cascaded continuous stirred tank heater process are used to present the superiority of the proposed method.