The support vector machine（ SVM） based on Mahalanobis distance（ MD） is widely used in classification of industrial product appearance defects because it has the characteristics of considering the overall distribution of samples

eliminating the coupling relationship between various features

and not related to dimension. Considering that the internal and inter-class covariance matrices of the total association variance matrix in MD have different effects on the classification results

different weight coefficients are set for the two

and then a weighted MD-SVM（ WMD-SVM） model is proposed. Trained sample set obtain optimal weight coefficients. Suitable kernel function is selected for different defect image features

and the proposed improved model is applied to each kernel function

which constitutes multi-core learning model. The fuzzy constraint theory is used to solve the weight of each kernel function in the model. Experimental results show that the improved method has better classification precision and real-time performance.