Monitoring nonstationary processes based on recursive cointegration analysis and elastic weight consolidation
arXiv:2101.08579
Abstract
This paper considers the problem of nonstationary process monitoring under frequently varying operating conditions. Traditional approaches generally misidentify the normal dynamic deviations as faults and thus lead to high false alarms. Besides, they generally consider single relatively steady operating condition and suffer from the catastrophic forgetting issue when learning successive operating conditions. In this paper, recursive cointegration analysis (RCA) is first proposed to distinguish the real faults from normal systems changes, where the model is updated once a new normal sample arrives and can adapt to slow change of cointegration relationship. Based on the long-term equilibrium information extracted by RCA, the remaining short-term dynamic information is monitored by recursive principal component analysis (RPCA). Thus a comprehensive monitoring framework is built. When the system enters a new operating condition, the RCA-RPCA model is rebuilt to deal with the new condition. Meanwhile, elastic weight consolidation (EWC) is employed to settle the `catastrophic forgetting' issue inherent in RPCA, where significant information of influential parameters is enhanced to avoid the abrupt performance degradation for similar modes. The effectiveness of the proposed method is illustrated by a practical industrial system.
arXiv commentThis paper has been submitted to IEEE Transaction on Cybernetics for potential publication
References in corpus (6)
- Overcoming catastrophic forgetting in neural networks
- Progress & Compress: A scalable framework for continual learning
- Alleviating catastrophic forgetting using context-dependent gating and synaptic stabilization
- An improved mixture of probabilistic PCA for nonlinear data-driven process monitoring
- Process monitoring based on orthogonal locality preserving projection with maximum likelihood estimation
- Monitoring multimode processes: a modified PCA algorithm with continual learning ability