Research On Key Technologies Of Soft Sensor For Product Quality Index Of Dynamic Time-varying Chemical Processes Based On Improved LSSVM

With the development of modern industry and technology,the on-line estimation of product quality indicators for complex chemical processes based on the soft sensor modeling method is becoming a research hotspot in the field of process control.Traditional least squares support vector machines,LSSVM(LSSVM)method has realized the on-line estimation of product quality index through static assumption.However,the actual chemical process has inertia,the system output can not reflect the input in real-time without distortion,and the inputoutput relationship is always in an unstable state.Therefore,the actual chemical process has dynamic and time-varying characteristics,which leads to the low predicted accuracy and strong limitations of the soft sensor model.Meanwhile,LSSVM does not fully consider the interference of external conditions such as equipment deterioration and production environment change,which affects the predicted accuracy of the soft sensor model that has been put into operation.To improve the predicted accuracy of the soft sensor model for chemical process,based on the in-depth analysis of dynamic and time-varying characteristics of complex chemical process data and improved LSSVM method,the dissertation first proposes a multioutput soft sensor modeling method based on improved ensemble learning.Next,a dynamic soft sensor modeling method based on a discount weighted ARMA model is developed.And then,a semi-supervised dynamic soft sensor modeling method based on a discount weighted MA model is proposed.Finally,a deterioration evaluation and model updating method of soft sensor model for the time-varying chemical process is investigated.The main work of this paper is as follows:Aiming at the problem that the traditional single soft sensor model is difficult to describe complex data relationships,a multi-output soft sensor modeling method based on improved ensemble learning is proposed.The existing soft sensor modeling methods ignore the internal relationship between process variables and different product quality indicators,so it is difficult to reflect the influence of process variables on different product quality indicators.In this paper,an auxiliary variable selection method based on cosine similarity is designed,and then an improved bagging integrated learning strategy is constructed to realize a more reasonable multioutput soft sensor modeling that can reflect the process variables and product quality laws.Besides,a multi-index soft sensor model fusion method based on a weighted integration strategy is proposed for the overall evaluation of product quality.Finally,the method is verified on the industrial data set of the sulfur recovery unit and the simulation data set of the penicillin fermentation unit.The results show that the proposed method has higher predicted accuracy than the traditional ensemble learning method.Aiming at the problem that the existing dynamic soft sensor modeling methods are limited to the fixed-length auxiliary variable historical data,and the multi-rate data information is not fully utilized,a dynamic soft sensor modeling method based on discount weighted ARMA model is proposed.This method uses autoregressive moving average(ARMA)to add the last batch output as the current batch input to the modeling sample data set,which can effectively capture the nonfixed length process variable data information,and establish a dynamic weighted input model to realize the transformation from time dynamic modeling to spatial static modeling.Furthermore,the weight calculation method of historical data fusion of process variables based on the discount method is designed,which effectively enhances the credibility of process variable data with a short time span from the current batch output.Finally,the simulation experiments are carried out with the actual data set of the distillation process and the simulation data set of the continuous stirred tank reactor.The results show that the soft sensor model established by the proposed method has better predicted accuracy.To solve the problem that existing soft sensor modeling methods are difficult to effectively use unlabeled samples that contain a large number of process dynamic information,a semisupervised dynamic soft sensor modeling method based on a discount weighted MA model is proposed.The semi-supervised clustering of unlabeled samples is effectively realized by semisupervised clustering method with constrained cosine similarity,and the influence of abnormal data on semi-supervised clustering data set is reduced.By combining the dynamic weighted input model with the moving average(MA)model,the weighted fusion of nonfixed length semi-supervised clustering data is realized,and the multi-rate problem of semi-supervised clustering sample data is solved.Besides,the semi-supervised clustering data fusion weight calculation method based on the discount method enhances the influence of semi-supervised clustering data with a short time span from the leading variable data on the modeling auxiliary variable data.This method improves the processing ability of unlabeled samples and expands the application range of unlabeled samples.Finally,the simulation experiments were carried out with the actual data set of the distillation process and the simulation data set of the continuous stirred tank reactor.The results show that the soft sensor model established by the proposed method has better predicted accuracy.To solve the problem that the existing evaluation methods of soft sensor model deterioration are easy to be affected by abnormal data,and the predicted accuracy of the updated soft sensor model is limited,a method for deterioration evaluation and model updating of soft sensor model for time-varying chemical process is proposed.Based on the statistical hypothesis testing theory,according to the deviation data of the predicted performance evaluation index between the soft sensor model and the initial soft sensor model after the operation,a statistic that can test whether the predicted deviation has significant change is constructed.Combined with the confidence interval threshold,the deterioration evaluation of the soft sensor model is realized.Besides,the relatively new sample data of chemical process variables are obtained by the moving window method.Combined with the proposed cosine similarity calculation method with constraints,the adaptive updating of auxiliary variables and data of the soft sensor model is realized,and the soft sensor model is rebuilt.Finally,the simulation experiments were carried out using the actual data set of debutanizers and the simulation data set of the continuous stirred tank reactor.The experimental results show that the proposed method can effectively reduce the influence of chemical process time-varying characteristics on the predicted accuracy of soft sensor model,realize the accurate evaluation of the deterioration of soft sensor model for complex chemical process,reduce the misevaluation rate,improve the predicted accuracy of the updated soft sensor model,and greatly enhance the application prospect in the actual industrial process.