Adaptive soft sensors for quality prediction under the framework of Bayesian network

Soft sensor is widely used to predict quality-relevant variables which are usually hard to measure timely. Due to model degradation, it is necessary to construct an adaptive model to follow changes of the process. Adaptive models-moving windows (MW), time difference (TD), and locally weighted regression (LWR) under the framework of Bayesian network (BN) are proposed in this paper. BN shows great superiorities over other traditional methods, especially in dealing with missing data and the ability of learning causality. Furthermore, the acquisition of variances in BN makes it possible to perform fault detection, on the basis of 3-sigma criterion. A debutanizer column and CO2 absorption column are provided as two industrial examples to validate the effectiveness of our proposed techniques. In a debutanizer column, RMSE of MW-BN is decreased by 40% in comparison to MW-PLS. In a CO2 absorption column, the largest absolute prediction error of TD-BN is reduced by approximate 7% when compared with that of TD-PLS. Furthermore, about 38% improvements of prediction precision can be achieved in LW-BN in contrast to LW-PLS.