A soft sensor is an empirical model, which estimates variables that is infeasible to measure on-line from other correlated variables. Because constructing a soft sensor is a process of data based empirical modeling, the homogeneity in the training data set is very important. If a process experiences a wide operation range or lasting significant disturbances, the data homogeneity is damaged and frequently results in several sub-clusters, which causes the prediction power of a soft sensor to be degraded. In this paper, we proposed a modeling procedure,that involves classification of training data and subclass modeling. PLS and NLPLS are adopted selectively as the subclass modeling algorithm and wavelet coefficients thresholding is used to remove noises contained in signals without severe distortion of the signals. Also, weighted X variables based on Variable Importance to Projection(VIP) through exploratory PLS are used to enhance the performance of the subclass models. The proposed methodology has been illustrated using an application to the development of a soft sensor for composition estimation of a binary distillation column simulated with HYSYS. The soft sensor based on the proposed scheme has shown better performance and robustness.