Soft Sensor Methods Study On Product Quality Based On Variable Selection

In industrial production,some product qualities are hard to measure directly,and the development of soft sensor technique has efficiently solved such problems.The prediction of product quality is a necessary procedure to ensure the maximum of product benefits,and the secondary variable selection is one of the key techniques of soft sensor modeling.Due to the development of detection technology,more and more industrial process data are collected,so it is crucial to study how to preform variable selection by removing uninformative and redundant variables and enhance the robustness and prediction accuracy of the model.From the perspective of variable selection,this thesis studies how to improve the prediction accuracy of industrial products.Aiming at the uninformative and redundant variables in near infrared spectrum data and the drawback of elastic net in processing high-related variables,this thesis proposed an elastic net with regression coefficients method for variable selection.This method preforms variable selection based on the coefficients of elastic net,and gets a sparse model.It can not only improve the prediction accuracy,but also can enhance the interpretability and robustness of the model.Two case studies have demonstrated its efficiency.Based on the mutual information,this thesis proposed a variable selection method for quality prediction.This method firstly preforms the linearity test between process data and quality data,and then calculates the importance order of process variables.Start from the most important variable,the variables are added into partial lest squares model one by one.Finally,the best variable subset is obtained according to the root mean square error criterion.This method can improve the prediction accuracy and model stability,and its superiority is verified by two case studies.For dynamic and time-varying industrial process,a mutual information based dynamic moving window least square support vector machine algorithm is proposed.After preforming variable selection based on mutual information,through introducing moving window technique,this method real-time updates the training data by incremental algorithm and decremental algorithm,and builds online product prediction model.Finally,a distillation process case study demonstrates that the proposed method can decrease the computation load and get a high prediction accuracy.