Few-Shot Learning with Contrast Prototype Network for Fault Diagnosis

The identification of mechanical faults is crucial for guaranteeing the secure functioning of machinery. While deep learning techniques for fault diagnosis have gained popularity, their majority necessitate a substantial quantity of labeled data for their training. As a result, this research suggests few-shot learning of contrast prototype (CPNet)-based techniques for problem diagnosis in small sample sizes. First, CPNet acquires multimodal fault features by constructing a multidimensional convolutional network. Furthermore, the network backbone (Backbone) is improved with a multi-head attention mechanism to enhance the model's feature extraction performance. Secondly, supervised contrast learning and multimodal feature similarity learning are used to extract intra-class and inter-modal fault features to facilitate the model to recognize faults in small-sample fault diagnosis tasks. Finally, fault classification is achieved by computing multimodal feature prototypes. The suggested approach is verified using a bearing dataset, and the outcomes demonstrate that it produces competitive outcomes on small samples.