Abstract: A regression method is proposed to improve the LS-SVM (least square-support vector machine) model. The sparseness of LS-SVM is thus obtained from the regression model to increase greatly the computation speed by way of removing most of the sample points in accordance to the Euclidian distances between input variables and sample points. The proposed method has been applied to the fermentation process to develop a soft sensing model so as to estimate the product's concentration on-line in penicillin fermentation process. Simulation results showed that the proposed method can provide a new useful approach to the real-time monitoring of quality variables which are hard to measure on-line in fermentation processes.