Quality modeling and monitoring for the linear-nonlinear-coexistence process

Abstract Linear and nonlinear relationships may exist simultaneously across process variables and quality variables. If only the linear model is established, the nonlinear structure may be neglected. If only the nonlinear model is constructed, the model accuracy and monitoring performance may be degraded. Thus, the quality monitoring method considering both linear and nonlinear relationships needs to be presented. In this paper, a serial ridge regression (SRR) method is proposed for quality monitoring in the linear-nonlinear-coexistence process. Firstly, linear features are extracted to construct the linear-quality-feature subspace, and the remaining information constitutes the complementary feature subspace. Then, the nonlinear-quality-features are further extracted from the complementary feature subspace via kernel-based strategy. Thereafter, in order to obtain more direct and clear monitoring results, the quality monitoring index is developed based on Bayesian inference. Case studies in a numerical simulation, continuous stirred tank reactor (CSTR) process and TE process demonstrate that the SRR-based method significantly outperforms partial least squares (PLS), ridge regression (RR) and kernel ridge regression (KRR)-based methods, in terms of higher fault detection rates, lower false alarm rates and better fault sensitivity. It helps the operator to detect faults earlier and avoid unnecessary downtime and maintenance.