In order to solve the problem of traditional methods having low accuracy in recognizing rolling bearings' faults and to reduce time in building a training model, this paper puts forward a method of recognizing faults based on wavelet packet transformation and PCA-PSO-MSVM. First of all, we extracted the energy values after all kinds of faulty signals have been transformed through wavelet packet, and combined the index of the signals' time-domain characteristics to form an eigenvalue matrix to be used as the characteristics of SVM sample training. Then by using PCA (principal component analysis) method, we eliminated the redundant characteristics to increase the efficiency of the model structure. Afterwards, through MSVM algorithm and one-to-one strategy, we carried out final multi-classification recognition of the mixed faults of rolling bearings, with PSO (particle swarm optimization) algorithm also adopted for optimizing SVM parameters to improve the accuracy of classifications. The experiment adopted the data of rolling bearings provided by Case Western Reserve University. As shown in the experimental results, this method has high accuracy and efficiency.