To overcome the difficulties in recognizing the compound faults of mechanical equipment, a novel mechanical fault diagnosis model based on multi wavelet packet, neighborhood rough set and multi-class support vector machine (SVM) is proposed. In the model, the raw signals are decomposed via multi wavelet packet, and the features in time domain are extracted from the raw signals and each decomposed signal to construct the original features. Then, the sensitive features are selected by neighborhood rough set method and input into the multi-class SVM to automatically identify different conditions of mechanical equipment. The model is applied to fault diagnosis of rolling bearings in electric locomotive. The application results demonstrate that multi wavelet packet enables to extract more abundant fault information and obtain higher classification accuracy than traditional wavelet packet; neighborhood rough set enables to select sensitive ones from a large number of original features to rapidly and more correctly diagnosis the mechanical faults. Therefore, this model possesses significant advantage in diagnosis of compound faults.