Abstract:Using deep learning technique to automatically and accurately identify the incipient fault of rolling bearing, especially the fault position, classification and severity degree, is a research hotspot in current fault diagnosis field. The traditional fault diagnosis method excessively relies on the manual feature extraction by the engineers with prior knowledge, which is difficult to effectively extract incipient fault features. In this paper, a novel improved CNNsSVM method is proposed and used for the rapid intelligent fault diagnosis of motor rolling bearing. This method adopts the combination of 1×1 transitional convolution layer and global average pooling layer to replace the fully connected network layer structure of traditional CNN, which effectively reduces the number of training parameters of CNN. In test stage, the method uses SVM to replace the Softmax classifier, which further improves the diagnosis accuracy. The proposed method was applied to the fault experiment data of the motor support rolling bearing, and the method was compared and verified with traditional intelligent diagnosis methods. The results show that the accuracy of fault identification of the improved CNNsSVM algorithm reaches up to 9986%, and the proposed method has good migration generalization ability under different load conditions and possesses the feasibility for practical engineering application. The fault diagnosis accuracy and test time of the method is obviously better than other intelligent algorithms.