The traditional deep learning-based bearing fault diagnosis approaches assume that the training and test data follow the same distribution. This assumption, however, is not always true for the bearing data collected in practical scenarios, leading to significant decline to fault diagnosis performance. In order to satisfy this assumption, the transfer learning concept is introduced in deep learning by transferring the knowledge learned from other data or models. Due to the excellent capability of feature learning and domain transfer, deep transfer learning methods have gained widespread attention in bearing fault diagnosis in recent years. This article presents a comprehensive review of the development of deep transfer learning-based bearing fault diagnosis approaches since 2016. In this review, a novel taxonomy of deep transfer learning-based bearing fault diagnosis methods is proposed from the perspective of target domain data properties divided by labels, machines, and faults. By covering the whole life cycle of deep transfer learning-based fault diagnosis and discussing the research challenges and opportunities, this review provides a systematic guideline for researchers and practitioners to efficiently identify suitable deep transfer learning models based on the actual problems encountered in bearing fault diagnosis.

In bearing fault diagnosis, the faulty data are generally limited due to the high cost of fault signal collection. Considering the excessive parameters in the traditional convolutional neural network (CNN), such a limited data issue can cause overfitting problem during the model training, eventually resulting in poor fault diagnosis performance. To resolve the overfitting issue and elevate the diagnostic accuracy of the conventional methods, a novel fault diagnosis method based on local binary temporal convolutional neural network (LBTCNN) is proposed in this paper. In the proposed LBTCNN, a novel temporal module with dilated causal convolution for deep feature extraction is proposed to increase model depth under limited model parameters, and a local binary convolution (LBC) layer is adopted to reduce the computational parameters. To evaluate the effectiveness of the proposed method, several experiments under different scenarios such as limited samples and different noise levels are conducted on two datasets, including the rolling bearing accelerated life test dataset of Xi’an Jiaotong University and Changxing Sumyoung Technology, and the motor bearing dataset of Case Western Reserve University. The comparison results demonstrate that the LBTCNN method is superior over six other prominent fault diagnosis approaches under different bearing operation stages, different training samples and different signal-to-noise ratios.

In the electroencephalography (EEG) based cross-subject motor imagery (MI) classification task, the device and subject problems can cause the time-related data distribution shift problem. In a single-source to single-target (STS) MI classification task, such a shift problem will certainly provoke an increase in the overall data distribution difference between the source and target domains, giving rise to poor classification accuracy. In this paper, a novel multi-subdomain adaptation method (MSDAN) is proposed to solve the shift problem and improve the classification accuracy of the traditional approaches. In the proposed MSDAN, the adaptation losses in both class-related and time-related subdomains (that are divided by different data labels and session labels) are obtained by measuring the distribution differences between the source and target subdomains. Then, the adaptation and classification losses in the loss function of MSDAN are minimized concurrently. To illustrate the application value of the proposed method, our method is applied to solve the STS MI classification task about data analysis with respect to the brain-computer interface (BCI) competition III-IVa dataset. The resultant experiment results demonstrate that compared with other well-known domain adaptation and deep learning methods, the proposed method is capable of solving the time-related data distribution problem at higher classification accuracy.