Fault Diagnosis Method for Small Sample Bearing Based on Transfer Learning

Aiming at the problem of training sample shortage in practice, an improved transfer learning network is proposed which transfers the fault diagnosis knowledge learned by the model in source domain to target domain, and is used in small sample bearing fault diagnosis. The global average pooling layer is used instead of the fully connected layer in the convolutional neural network to reduce training network parameters. The fine-tuning transfer learning method is adopted to train the network with sufficient source domain samples, which avoids over-fitting caused by insufficient data. After transferring the network structure and parameters to the target domain, the deeper network parameters are fine-tuned to make the network adapt to the data distribution of the target domain samples. The fault diagnosis experiments are carried out on Case Western Reserve University bearing data set and the laboratory bearing data set. The results show that the transfer learning network enables to achieve a classification accuracy of 92.25% taking only 1% target domain training samples, in the case of cross-working conditions and cross-bearing types. This proposed method completes the transfer learning classification taking small samples, and provides a reference for research and application of transfer learning theory in bearing fault diagnosis. © 2021, Editorial Office of Journal of Xi'an Jiaotong University. All right reserved.