Integrating Novel Convolutional Neural Network with Prototype Learning for Rotating Machinery Fault Diagnosis under Small Sample

Abstract Rotating machinery serves as essential components of modern industrial systems, demanding critical research into fault diagnosis methods. However, insufficient training data, caused by high labeling costs and limited data acquisition, leads to overfitting in deep learning models. Additionally, discrepancies between fault signals and training data under complex conditions hinder the extraction of critical features, degrading diagnostic performance. To address these issues, we propose a novel CNN with prototype learning for small sample fault diagnosis under complex conditions. The feature extractor incorporates convolutional blocks with feature restoration layers, which employ multi-kernel channel attention to enhance the extraction and integration of task-related features, thereby mitigating the adverse effects caused by data distribution discrepancies. Furthermore, discriminant boundary loss and clustering loss are integrated with the prototype learning loss function, making sample features in the feature space more distinguishable. Experiments on two datasets demonstrate superior diagnostic accuracy and clearer feature visualization under small sample and complex conditions, validating the method's generalizability and effectiveness.