Since the non-stationary of vibration signals cannot be fully described by the stationary autoregression model, a feature extraction approach based on wavelet packet decomposition(WPD) and autoregressive(AR) model is proposed, and then the feature vectors are extracted to accurately reflect the running state of rolling bearing. Firstly, the non-stationary signals generated by rolling bearing vibration are decomposed into some coefficients by wavelet packet transformation. Then, the coefficients are modeled as AR model and the parameters of AR model are used as the feature vectors. Finally, fault patterns are recognized by the feature vectors using support vector machine (SVM) classifier, consequently the intelligent fault diagnosis is realized. The simulation results show the effectiveness of the proposed method.