Variable predictive model-based class discriminate (VPMCD) classification method was built on homoscedastic regression model. When the regression model is heteroscedastic, it will lower the prediction accuracy. So a WLS -variable predictive mode-based class discriminate (WVPMCD) pattern recognition method was presented. Its parameter estimation approach uses the weighted least square method to replace ordinary least square method to eliminate homoscedasticity, thus raising the accuracy of pattern recognition. In this paper, LCD (Local characteristic-scale decomposition) approach was adopted to decompose the roller bearing vibration signal. Then, the singular values are abstracted from the component matrix and formed into fault feature vector which will act as the input in WVPMCD. The analysis results from the roller bearing vibration signals of normal, roller fault, inner race fault and outer race fault demonstrate that WVPMCD has a higher recognition rate when the regression model is heteroscedastic．