Ensemble Adaptive Kernel PLS Soft Sensing Modeling Method And Its Application

With the rapid development and improvement of industrial progress,conventional detection technologies cannot meet all the requirements from control,thus more and more attention has been turned to the soft sensing technology.As is known,soft sensor modeling is the core of the soft sensing technology,where varieties of modeling methods have been proposed and discussed.Among these modeling methods,Partial Least Squares(PLS)is a widely used one because of its advantages in effectively handling small samples,multiple noises and serious collinear variations of variables.However,as a traditional linear modeling method,PLS's nonlinearly modeling ability is limited.The industrial data tend to be strongly nonlinear,therefore it cannot be modeled directly using PLS.In summary,to improve the PLS nonlinear modeling capabilities is one of the hot topics today,and it is the main focus of this thesis.The thesis was supported by the National Natural Science Foundation of China and the Natural Science Foundation of Zhejiang Province.Its main research and achievements are summarized as follows:(1)The combination of kernel function and local weighted algorithm can be considered as a dual guarantee for the fitting of nonlinear data.At the same time,on the basis of kernel PLS algorithm,this thesis proposed an adaptive selection mechanism of kernel function based on particle swarm optimization,so that the mapping relationship between kernel function and training samples is more consistent with data distribution characteristics,so as to further improve the prediction accuracy of the model.(2)An ensemble adaptive kernel PLS algorithm based on K-means is proposed for the nonlinear data without obvious timing characteristics.In the process of model training,the K-means algorithm is used to cluster the data in space.Then select the optimal kernel function and build the corresponding sub models for each subset by adaptive kernel PLS(AKPLS).In the process of testing,this thesis puts forward the idea of calculating the weights first and then deciding the prediction effect of each sub model for the output of test data,so that the steps of useless prediction of low matching sub models can be avoided,so as to reduce the computation cost of models.In the weight calculation step,the concept of absolute neighborhood is proposed,which improves the accuracy and operation efficiency.(3)An ensemble adaptive kernel PLS algorithm based on moving windows is proposed for time-series nonlinear data.The similarity of linked data can be used to find out the moment of state mutation through the moving window,and divide the data set.The model pruning technique is added to delete redundant submodels and simplify the model structure.In the weight calculation step,the calculation of weight values considers the influence of two factors in the time domain and space domain both,which improves the generalization ability of the model.(4)The proposed two nonlinear ensemble PLS algorithms are applied to a coking system,and the temperature prediction model of the start line is established.Compared with traditional linear PLS,total kernel PLS and ensemble PLS,the comparison results show that the two improved algorithms proposed in this thesis have higher precision and better generalization performance for nonlinear data.At the same time,the feasibility and effectiveness of the two ensemble algorithms used in the practical industry are proved.