Dynamic Soft Sensor Development Based On Gaussian Mixture Regression

In industrial processes,many key variables are difficult to measure online.Soft sensor technology has become a new method to solve this problem.At present,many models are based on the assumption that processes are in steady state.So,many soft sensor models are static models.However,production processes are often in dynamic state due to changes of production processes,physical structures and environmental factors.Therefore,it is necessary to develop dynamic soft sensor modeling methods to overcome disadvantages of static models.In this paper,dynamic soft sensor models are established by using the Gaussian process regression(GPR)and Gaussian mixture regression(GMR)on a simulated penicillin process and an industrial erythromycin fermentation process.Specific works are as follows:(1)Aiming at the problem of low precision and poor robustness of static soft sensor models,a dynamic soft sensor modeling method based on the multi-criteria and GPR is proposed.In this paper,a novel the optimization strategy of the GPR based dynamic soft sensor model is proposed,which lays the foundation of the order of the model.Finally,the proposed model was used to estimate the concentration of erythromycin fermentation process.The results show that the dynamic soft sensor modeling method based on GPR can achieve high prediction of biomass concentration,and the prediction have small confidence intervals.(2)For the limitation of dynamic soft sensor model based on single GPR in fermentation process,a dynamic soft sensor model based on GMR is proposed in this work.The model has two important parameters,namely,the number of Gaussian elements and the order of the model.In order to obtain the optimized soft sensor model,a novel iterative strategy is proposed to optimize the two structural parameters.Finally,the proposed dynamic GMR soft sensor was applied to a simulated penicillin process and an industrial erythromycin fermentation process to estimate biomass concentrations.For comparisons,the existing dynamic GPR soft sensor was also studied.The results show that the dynamic GMR soft sensor has high prediction accuracy and is more suitable to model dynamic multiphase / multimodal fermentation process.(3)In order to reduce update frequency moving window algorithms and GMR based soft sensor model(MW-GMR),a recursive GMR modeling method based on performance evaluation(MPA-GMR)is proposed.Firstly,the initial confidence limits of the model are generated automatically.The predicted root mean square error(RMSE)is used as the index of model performance.Secondly,model calibration is selectively activated and the confidence limits are updated with the index under some conditions.Finally,the developed model was used to estimate the biomass concentration during penicillin simulation and industrial erythromycin fermentation.The results show that the developed model greatly improves the computational efficiency,and the loss of prediction accuracy can be neglected.