Adaptive monitoring for multimode nonstationary processes using cointegration analysis and probabilistic slow feature analysis

The condition monitoring of nonlinear, nonstationary and multimode processes is a difficult problem. Traditional multimode process monitoring methods generally assume that data from all potential modes are available, yet new modes may appear continuously in practice. This paper investigates an intelligent adaptive monitoring method for multimode nonstationary processes, which can deal with the appearance of new modes with ease. A comprehensive framework is proposed to decompose feature subspaces. First, long-term equilibrium features are extracted by adaptive cointegration analysis (ACA) to identify the mode, without using any prior mode information intelligently for online applications. Then, recursive attention probabilistic slow feature analysis integrated with elastic weight consolidation (RAttPSFA-EWC) is investigated to deal with the remaining dynamic information and extract dynamic and static slow features to maintain continual learning for multimodes. Once anew mode is detected automatically, the previously learned knowledge is consolidated while extracting new features, which is beneficial to enhancing the performance of similar modes. The proposed ACA-RAttPSFA-EWC acts as an online adaptive method by parameter updates with incoming normal data. Furthermore, several advanced methods are compared to demonstrate the strengths of ACA-RAttPSFA-EWC, and the proposed method is validated to be effective using a numerical case and a practical system.