Abstract In the process of industrial production, bearing fault detection has always been a hot issudza20000528@163.comsolved. At present, the problem of less fault data samples in the field of fault detection has caused great trouble to the research of deep learning. In the application of industrial fault detection, which is difficult to obtain massive data, it is easy to lead to the lack of fitting of neural network training and many generalization problems. To solve the above problems, this paper proposes an improved and more efficient method of few-shot supervised learning, which is called the Temporal Classification Prototype Network (TCPN). This model is designed to maintain both training efficacy and generalization capabilities under conditions of data scarcity. Initially, Fourier transform is employed to accentuate the frequency domain characteristics of the fault section in the bearing signal before it is input into the model, thereby enabling the subsequent model to concentrate on distinguishing between normal and fault signals. Subsequently, discrete data sample points are transformed into points within the feature space via our Enhanced Temporal Convolutional Network(ETCN). In our investigation, we utilize the features of the support set as anchors within the feature space and employ similarity measures as the basis for classification, thus developing a more effective comparative learning classifier known as the ContractSim Classifier (CSC). Within the CSC, the model learns the data features of the query set, which are then back-propagated to refine our model. The proposed TCPN model has been evaluated across four standard bearing datasets, corroborating its few-shot learning proficiency through k-shot experiments. In comparative model experiments, our TCPN outperforms baseline models, while the ablation study confirms the rationality and robustness of our module integration.