Research On Soft Sensor Modeling Based On Slow Feature Analysis

Soft-sensor technology mainly constructs an inference model through easy-to-measure process variables(auxiliary variables)to realize online estimation of unmeasured variables(leading variables),which has been widely used in industry.However,with the increasing of the complexity of process technology and optimization of advanced control systems,the amount of process data is also increasing exponentially.These require higher accuracy and reliability for the soft-sensor technology and make the extracting of important feature information from the complex processes becomes more important.In this paper,slow feature analysis(SFA)is used as the basic algorithm to extract feature data from process data.At the same time,the nonlinearity and time delay in the industrial process data are considered,and the soft sensor modeling method is improved.The main research results are as follows.(1)Aiming at the feature extraction of industrial process data,a regression modeling method based on slow feature analysis is proposed.Firstly,the slowest feature analysis algorithm is used to extract some of the slowest components of the process data as the essential features of the process.Secondly,a least-squares regression model is built based on the essential features.Finally,the TE process data is selected for simulation experiments.The compared results of the principal regression and partial least squares regression modeling methods show that the slow feature analysis algorithm is effective,and the regression model based on this algorithm has higher prediction accuracy.(2)Considering the nonlinearity widely existed in most practical industrial processes,a method of Gaussian process regression soft sensor modeling based on nonlinear slow feature analysis is proposed.Firstly,the linear SFA is improved by the nonlinear slow feature analysis of the second-order polynomial expansion form,so as to extract the nonlinear features in the process.Secondly,a Gaussian process regression(GPR)model is established by using the extracted nonlinear features.Finally,through the data simulation and analysis of the penicillin fermentation process,the validity of the nonlinear slow feature analysis method is verified,and the modeling accuracy is improved.(3)Further considering the inconsistency of modeling data due to the time delay in the actual industrial process,and improving the performance of the second-order polynomial expansion form of nonlinear slow feature analysis algorithm,a method of GPR modeling based on kernel slow feature analysis and time delay estimation is proposed.Firstly,the historical data is mined by fuzzy curve analysis to obtain the optimal time-delay parameter estimation,and the modeling data is reconstructed.Then,the nonlinear feature extraction is performed on the reconstructed by data using the kernel slow feature analysis.Based on the extracted features,establish a GPR model to achieve the prediction of the dominant variables.Through the simulation experiment of the debutanizer process,the effectiveness and precision of the proposed method are verified.