An Improved Data-driven Soft Sensor Modeling Algorithm Based on Twin Support Vector Regression for Sugar Cane Crystallization

Abstract

According to the problem that some key parameters, such as mother liquor supersaturation, mother liquor purity, crystal content and crystal size distribution, cannot be measured on-line during sugar cane crystallization, an improved data-driven soft sensor modeling algorithm based on twin support vector regression is proposed. Following improvements are taken based on traditional data-driven model. The complexity of data-driven model is decreased by adding a regularization term, which can transform empirical risk into structural risk. Computational speed is increased and computational time is decreased efficiently by modifying the size of kernel function matrix. Different punishment weight is given to sample sets according to their own importance, which can increase the algorithm's generalization ability and avoid over-fitting problems to a certain degree. Experimental results show that compared with traditional data-driven soft sensor modeling, this improved algorithm has better prediction result and less prediction error than traditional data-driven modeling method.