Adaptive Multi-model Modeling And Correction Based On Gaussian Process Regression

Due to the limitation of detection technology or operation condition, some important process variables in complex industrial process are difficult to mearsure online directly. To overcome this difficulty, soft sensor has gained much attention in industrial process. In this paper, based on current soft sersor, gaussian process regression(GPR) is applied as modeling method. Aiming to achieve constructing adaptive models and correcting built soft sensor model, more researches have been made.Firstly, reliable quality prediction of chemical processes often encounters various challenges including process nonlinearity, multiple operating phases and different local dynamics. To address these problems, an online adaptive multi-model soft sensor is proposed. The gaussian mixture model(GMM) is firstly introduced to distinguish the data from different operating phases. Then, a just-in-time learning(JITL) strategy is applied to update the local GPR model. Whenever a new sample is available, according to the posterior probabilities of the sample belonging to the different operating phases, the predictions of the local GPR models are combined to obtain the desired global output. A Tennessee Eastman(TE) chemical process is employed to demonstrate the feasibility and effectiveness of the proposed approach. The results show the proposed approach has higher prediction accuracy and better generation ability.Secondly, to make the moving window(MW) strategy work more efficient, JITL algorithm is employed to enhance the performace of conventional MW approach. However, the MWGPR based models are always static. In consideration of industrial dynamic characteritic, an ARX model structure is proposed.The GMM is firstly introduced to separate the data according to different operating modes. Then the MWGPR strategy is applied to identify the ARX model. Dual-updating and JITL are applied to make the identified dynamic model to effectively track process dynamics. A simulation of a continuous fermentation and a pilot scale experiment are presented to demonstrate the effectiveness of the proposed method. The results indicate that the proposed method can update current model and capture process dynamics effectively.Finally, non-normal noise can influence the quality prediction accuracy of soft sensor models. To address this problem, a GPR based error gaussian mixture model(EGMM) soft sensor is proposed. First, appropriate variables should be selected to form the error data and the optimal number of Gaussian component using Bayesian informance is determined. Then conditional error mean with respect to the new sample is computed using the EGMM model to compensate the prediction output so that more accurate prediction can be achieved. Through a numerical simulation and a soft sensor of predicting the H2 S concentrations of sulfur recovery unit(SRU), the feasibility and effectiveness of the proposed approach is demonstrated.