Prediction of Product Quality Based on Multiple Kernel Learning Least Square Support Vector Machine Regression

Some important product quality variables in the distillation column are difficult to be detected and predicted online. To solve this problem, a novel method of soft sensing based on multiple kernel learning least square support vector machine regression (MKL-LS-SVM) is proposed. To improve the generation of the soft sensor model, polynomial kernels, radial basis function kernels, sigmoid kernels are used in the multiple kernel learning LS-SVM algorithm. The application shows that multi-core instead of single kernel can improve the interpretability of decision function and obtain the better performance. By using such as root mean squared error as evaluation index to evaluate the learning algorithm in the nonlinear function data set and the continuous distillation simulation data set under noise disturbance. The simulation results show that this MKL-LS-SVM method can improve the prediction accuracy of soft sensing in the distillation process.