Research On Variable Selection For Soft Sensor Model

Soft sensor technique has been widely used in chemical engineering process.By using secondary variables as input and primary variable as output it can realize the online estimation of primary variable through the soft sensor technique.Although modelling is the core of a soft sensor,the soft sensor model will be effective only if the secondary variables closely related to the primary variable are selected.Variable selection is such a task to determine the best variable subset under a certain criterion which can describe the primary variable best.Assuming there are p candidate secondary variables,it will generate 2P-1 candidate models.Even if p is not very large,it will also fall into plight of the combination explosion.Therefore,it is necessary to study how to implement the variable selection in a quick and efficient way so that it can not only ensure the prediction performance as much as possible but also reduce the model complexity.To deal with the challenge,variable selection is studied in this thesis.The original contribution of this paper is summarized as follows:1.By combining Monte Carlo Uninformative Variable Elimination(MC-UVE)and Genetic Algorithm and Partial Least Squares(GA-PLS),this paper proposed the so-called MC-UVE-GA-PLS variable selection method,it can eliminate uninformative variables by firstly using the MC-UVE method;after the informative variables are selected,the genetic algorithm is further used to search the best subset of variables.At last,examples from UCI dataset are tested to show the good performance of the proposed method in a comparison with All-PLS model and GA-PLS model.2.Considering the variable selection is essentially a mathematical optimization problem,this paper proposed a nested MIQP-MLR variable selection method.It is based on the multiple linear regression(MLR)model and uses BIC criterion to formulate variable selection as a nested MIQP problem by introducing a set of binary variables.This method can simultaneously establish the prediction model and select the feature variables.Examples from UCI datasets validate the effectiveness and practical utility of the proposed method compared with the All-MLR,Stepwise-MLR and MINLP-MLR methods.3.Based on the MIQP-MLR variable selection method,a more robust Support Vector Regression(SVR)model is used.The standard MSE criterion is modified with the? insensitive function.The MILP-SVR variable selection method is proposed based on the modified MSE criterion.The proposed method does not require any specification of the variable number in the model,thus avoiding the adjustment of the penalty factor.In addition,the SVR model can use the kernel function technique to achieve the nonlinear function fitting.Computational results on UCI datasets validate the reliability and practical utility of the proposed method.4.The above-mentioned variable selection methods are applied to soft sensor model of M-phenylenediamine in an industrial distillation column.The secondary variables are selected firstly and then the corresponding soft sensor models are established.The simulation results of the actual industrial data show the reliability and high performance of the proposed methods.