Multiple-input Multiple-output Soft Sensors Based on Kpca and Mkls-svm for Quality Prediction in Atmospheric Distillation Column

In this paper a method based on kernel principal component analysis (KPCA) and mixed kernel least square support vector machine regression (MKLS-SVM) for online quality prediction in atmospheric distillation column is presented. Firstly, the KPCA is employed to reduce the input vector’s dimensions of the multiple-input multiple-output (MIMO) soft sensor and created the data set which required training the MKLS-SVM. Then, considering that the characteristics of kernels have great impacts on learning and predictive results of LS-SVM, LS-SVM based on mixed polynomial kernel and RBF kernel is adopted to build the soft sensor model. The parameters of the MKLS-SVM are adaptively selected by the real-cord multi-population genetic algorithm (GA) with elitist strategy, migration operator, self-adaptive mutation and crossover operator. The modeling process is described with emphasis on data preprocessing and variables selection. Finally, the simulation results show that the MIMO soft sensors have good abilities of model generalization and the predicted values are in good agreement with lab measurements.