AbstractIn recent years, many machine learning-based methods have emerged to detect faulty bearings. However, most of these methods may not be practical due to the need to collect a large number of fault samples for training. This paper developed a novel few-shot learning framework for the fault diagnosis of freight train rolling bearings. The proposed method has the capability to transfer the learning outcome from one bearing fault diagnosis model to another different but related task for which very limited training data are available. The authors established a single-wheelset platform to collect acceleration signals of different types of bearing faults. The authors preprocessed the data through data segmentation and frequency domain transformation, and divided the data into training and test sets according to a certain ratio. A one-dimensional convolutional neural network (1D-CNN) was established to automatically extract the features of the bearing vibration signals and classify the fault types. The authors implemented two few-shot learning methods through parameter fine-tuning and a conditional Wasserstein generative adversarial network (C-WGAN). A case study demonstrated the classification performance of the proposed models. The results showed that the diagnosis capability of the 1D-CNN in the frequency domain is significantly superior to that in the time domain. However, when the amount of data is small, the 1D-CNN model does not work. In contrast, the few-shot learning of bearing faults works well for both the fine-tuned CNN and C-WGAN models. Furthermore, the classification performance of the C-WGAN is better than that of the fine-tuned CNN when the training data are extremely limited.