Nonlinear process monitoring based on linear subspace and Bayesian inference

This paper proposes a novel linear subspace and Bayesian inference based monitoring method for nonlinear processes. Through the introduced linear subspace method, the original nonlinear space can be approximated by several linear subspaces, based on which different monitoring sub-models are developed. A new subspace contribution index is defined for variable selection in each subspace. Monitoring results are first generated in each subspace, and then transferred to fault probabilities by the Bayesian inference strategy. To make the final monitoring decision, subspace monitoring results are combined together with their fault probabilities. Additionally, a corresponding fault diagnosis method is also developed. To demonstrate the computationally efficiency of the proposed method, detailed comparisons of the algorithm complexity for different methods are provided. Case studies of a numerical example and the Tennessee Eastman (TE) benchmark process both show the efficiency of the proposed method.