Decentralized dynamic monitoring based on multi-block reorganized subspace integrated with Bayesian inference for plant-wide process

Abstract

Multi-block subspace monitoring technology has been widely used in the field of plant-wide process to solve the problem of high complexity between variables. However, recent multi-block partition approaches neither consider variable distribution nor comprehensively analyze quality-relevant fault. To solve this problem, a plant-wide process quality monitoring approach based on multi-block reorganized subspace integrated with Bayesian inference (MBRS-BI) is put forward in this paper. Firstly, in view of the distribution characteristics of data, process variables are divided into Gaussian and non-Gaussian subspaces by Jarque-Bera detection method. Subsequently, in the two subspaces, mutual information (MI) is further employed to determine quality-relevant and quality-irrelevant variables so as to obtain quality-relevant and quality-irrelevant sub-blocks. Dynamic principle component analysis (DPCA) and dynamic independent component analysis (DICA) methods are adopted to monitor Gaussian quality-irrelevant and non-Gaussian quality-irrelevant blocks, respectively. For the monitoring of Gaussian quality-relevant and non-Gaussian quality-relevant blocks, we propose total dynamic principle component regression (TDPCR) and total dynamic independent component regression methods (TDICR) methods to focus on analyzing the impact of faults on output quality. Finally, in order to further achieve plant-wide process dynamic monitoring, all corresponding statistical metrics of sub-blocks are reorganized through Bayesian inference (BI). The proposed method (MBRS-BI) is elaborated by the Tennessee-Eastman (TE) process.