FRDPC subspace construction integrated with Bayesian inference for efficient monitoring of dynamic chemical processes

Modern chemical processes are usually characterized by large-scale, complex correlation, and strong dynamics, and monitoring of such processes is imperative. This paper proposes a performance-driven fault-relevant dynamic principal component (FRDPC) subspace construction integrated with Bayesian inference method to achieve efficient monitoring for dynamic chemical processes. First, dynamic principal component analysis is employed to deal with both auto-correlation and cross-correlation among variables. Second, considering fault information has no definite mapping to a certain dynamic principal component (DPC) and the existence of non-beneficial DPCs may cause redundancy in the monitoring, an FRDPC subspace is constructed for each fault through the performance-driven DPC selection. Then new process measurements are examined in each FRDPC subspace as well as the residual subspace. The monitoring results in all subspaces are fused to a comprehensive index through Bayesian inference to provide an intuitive indication of the process status. Case studies on a numerical example and the Tennessee Eastman benchmark process indicate the efficiency.