Data-Driven Predictive Model Based on Locally Weighted Bayesian Gaussian Regression

In this paper, a data-driven modeling approach referred to as the ‘locally weighted Bayesian Gaussian regression (LWBGR)’ is proposed to develop soft sensor for nonlinear industrial process. The LWBGR handles nonlinearities by constructing localized models and accommodates process uncertainties by adopting the probabilistic way. An important feature of the LWBGR is that each localized model is a fully Bayesian Gaussian regression model where the mean and covariance are treated as random variables rather than deterministic parameters. By doing so the LWBGR can achieve enhanced capabilities in dealing with overfitting and numerical issues, which leads to higher predictive accuracy and more stable solution. The performance of the LWBGR is evaluated using a real-world industrial process, and the results demonstrate the advantages of the LWBGR over the deterministic and other probabilistic locally weighted models.