Research On The Improvement Of Gaussian Process Regression Online Soft Sensor Modeling

Most of the actual industrial process with nonlinear,time-varying and uncertainty,and traditional off-line soft sensor model can't to track the state parameter of those industrial processes all the time.In order to solve the above problems,the traditional off-line soft sensor method is usually improved to the online adaptive method,and the online soft sensor model parameters and database are processed and updated according to the real-time data,and to ensure that the soft sensor model can track the dynamic characteristics of the industrial process and with anti-interference ability,and improve the accuracy and performance of the soft sensor model at last.In order to realize the effective prediction and control of dominant variable in the process,this paper uses the Gaussian Process Regression to study the actual industrial process and obtain the corresponding soft sensor model at first,and then the algorithm whichi is dynamic update the model and test or compensation for the singular point is proposed to update the model parameters and process the data,and the validity of the proposed algorithm is proved by the experimental results at last.The main research contents of the paper are as follows:(1)The research of gaussian process regression algorithm in practical application.Firstly,the principle of Gaussian process regression algorithm is briefly analyzed.Then the regression algorithm is used to study the penicillin fermentation process and the corresponding soft sensor model is established.Compared with the traditional Least Squares Support Vector Machine,it is shown that the Gaussian process regression model has better predictive performance.(2)A method based on dynamic model updating algorithm is proposed,because practical industrial processes often with the time-varying characteristics.Firstly,an offline model based on Gaussian process regression(GPR)approach can be built using the training dataset,and each output value and its prediction error can be achieved.Then the prediction error will be analyzed,if the error mean is larger than the previously set threshold value,a global updating scheme is employed for GPR offline model by simultaneously updating the covariance matrix and the parameters of the covariance function.Otherwise,the local updating method which only updates the covariance matrix will be utilized.Afterwards,the final prediction results will be obtained from further compensation by an error Gaussian mixture model(EGMM).The effectiveness of the proposed method is verified through the simulation experiment on a real industrial sewage treatment process at last.(3)To handle the problem of the singular query sample which is encountered in the application of soft sensor for the practical industrial processes,an online soft sensor method considering the test and compensation for the singular point is proposed in this paper.Firstly,a soft sensor model can be built based on Gaussian process regression(GPR)approach using the training dataset.Then the pauta criterion is improved to test the new query samples with higher degree of accuracy.If the new query sample is determined as a singular point,an auxiliary model based method is provided to repair the singular point.Thereafter,the renewed query sample is predicted.Otherwise,the GPR soft sensor model can be used to estimate thenew query sample directly.It can ensure the validity of the new query sample point.The effectiveness of the proposed method is verified through the simulation experiment on a real sulfur recovery unit treatment process at last.