Research On Key Variable Regression Prediction Model Based On Complex Feature Mining

In complex industries,due to production technology or economic needs,there are many key industrial variables that cannot be directly measured,resulting in the inability to adjust the industrial process in time,thereby reducing product quality.This paper uses data-driven soft sensing methods to predict key variables in complex industries.Constructing a new nonlinear dimensionality reduction hybrid forecasting model for a variety of complex data features in industrial processes.The simulation and experiment are carried out with the experimental data of the three-phase flow device set up by Cranfield University,and the results are better than those of the existing mainstream methods.The specific research results of this paper are as follows:1)In view of the strong nonlinear relationship between industrial data,this paper applies the nonlinear support vector regression(SVR)model to the prediction of actual industrial key variables.Predicted results by fitting regression to real data,the prediction results of the model(PLS)and the ridge regression model(RR)are compared,the results show that the SVR model has better predictive ability for nonlinear data.2)In view of the existence of multiple variable indicators in the industrial process,the data is characterized by redundancy,this paper introduces the principal component analysis(PCA)method to extract information and linearly reduce the dimension of the data.The SVR models are compared,and the comparison results show that the PCA-SVR model condenses the data information and improves the prediction results of key variables.3)In view of the fact that the PCA model cannot mine low-dimensional structural information in high-dimensional data,this paper proposes to use the Local Linear Embedding Model(LLE)in manifold learning to perform nonlinear dimensionality reduction and feature extraction for variables,and construct nonlinear support based on limited embedding,vector regression model(LLE-SVR).The experimental results show that the LLE model enriches industrial data and extracts intrinsic information,which further improves the prediction results of the key variables of the SVR model.4)In view of the fact that PCA and LLE models cannot extract time series features in industrial data,this paper introduces a canonical variable analysis model(CVA)to extract the time series information between historical data and future data by building the correlation between them.Constructed nonlinear support vector regression model(CVA-SVR).And compared with the PCA-SVR model and the LLE-SVR model,the results show that the CVA-SVR model has better prediction accuracy and higher robustness.