In this paper, we present a weighted kernel Fisher criterion based on the feature extraction to improve the classification accuracy. The basic idea of the weighted kernel Fisher criterion is to bring the edged classes and points closer to the normal sample classes. The motivation of the work is to solve the problems on subclasses which may be overlapped when using the traditional clustering algorithm. The proposed method is applied to soft sensor modeling for the quality index in Bisphenol A production process. Numerical examples as well as an experiment are employed to demonstrate the effectiveness of the proposed method.