A Novel Incremental Gaussian Mixture Regression and Its Application for Time-varying Multimodal Process Quality Prediction

Data-driven soft sensor approach has been widely applied on real-time prediction and control of difficult-to-measure quality variables. Among these approaches, the Gaussian mixture regression (GMR) carries the potential of dealing with nonlinear and non-Gaussian industry problems, which has drawn increasing popularity and attentions in recent years. However, the fluctuation of raw materials, change of process environment, aging of instruments and other factors will have an effect on system performances over time. Hence, the lack of adaptive mechanism will make the GMR difficult to suit for time-varying processes and may cause large prediction errors. In order to model time-varying industrial processes and improve the adaptability of the conventional GMR, an adaptive soft sensor based on incremental Gaussian mixture regression (IGMR) is proposed in this paper. The incremental idea is integrated and an adaptive mechanism is added, which endow the proposed IGMR with the capability of adapting to new data in online environment. Compared to the moving window GMR (MWGMR) and the just-in-time learning GMR (JITLGMR), the feasibility and effectiveness of the proposed IGMR are verified both in a numerical simulation and a real-life industrial process experiment.