A Novel Streaming Variational Bayesian Supervised Factor Analysis for Industrial Adaptive Soft Sensor Modeling

Time-varying and state shifting are two of the primary process factors that deliver unsatisfactory prediction performance of soft sensors. Adaptive soft sensor is commonly the alternative practice to guarantee desirable predictive results. In this paper, combining with symmetric Kullback-Leibler (SKL) divergence, a novel streaming variational Bayesian supervised factor analysis (S-VBSFA) model is developed to scratch the process time-varying and state shifting features through online updating of the posterior of model parameters with the latest process data, which is much more efficient than traditional adaptive strategies with thorough local modeling. During the updating process, the SKL divergence is utilized to determine priors of the next variation Bayesian inference. As a result, the proposed algorithm not only reduces the time for adaptive model updating, but also improves the prediction accuracy. Besides, the tracking time delay for process variations are also reduced. An industrial process case has verified the superiority of the proposed approach.