Intelligent fault diagnosis for rolling bearing based on improved convolutional neural network

Convolutional neural network (CNN) has gradually become a research hotspot in the field of intelligent fault diagnosis. The full connection layer structure of the traditional CNN model has a problem of excessive training parameters, which makes the network training and testing time longer. To solve this problem, a novel method called improved CNN is proposed for the rapid fault diagnosis of rolling bearing. The proposed method improves the traditional CNN model structure by introducing the global average pooling technology to replace the fully connected network structure. The proposed method can effectively reduce the model training parameter quantity and test time. The deep learning training techniques are applied, such as data enhancement, dropout, and adaptive learning rate, etc. The proposed method is applied to the diagnosis of rolling bearing experimental data sets and compared with traditional intelligent diagnosis methods. The results show that the accuracy of fault identification of improved CNN algorithm is 99.04%. The accuracy of incipient fault diagnosis and test time consumption are obviously better than traditional CNN and other intelligent algorithms. The proposed method does not need any manual feature extraction operations on the raw data during the whole fault diagnosis process. The end-to-end algorithm structure of proposed method has good operability and versatility. © 2020, Nanjing Univ. of Aeronautics an Astronautics. All right reserved.