Soft sensors are computer programs and inferential estimators with an auxiliary role in predicting the variables that are impossible or difficult to obtain by available process measurements. This paper presents a soft sensor design using the Fuzzy C-Means clustering combined with the Recursive Finite Newton algorithm to train a nonlinear support vector regression (FCM-RFN-SVR) machine. In the proposed approach, samples are split into some partitions; then, by means of RFN-SVR, a local model for each partition is set. The soft sensor presented in this paper is designed to be implemented on a distillation column so that it can predict the bottom product benzene concentration. Compared to typical support vector regression (SVR), the results demonstrate that the proposed method is more powerful in improving the generalization ability of the soft sensor (smaller mean squared error (MSE) and larger coefficient of determination (R-2) for the testing data).