Multi-source Heterogeneous Information Fusion Prototype Network Based on Compact-Sparse Representation for Rotating Machinery Few-Shot Fault Diagnosis

Scarce and singular fault monitoring data hinder the engineering application and generalization of deep learning-based fault diagnosis models to a certain extent. To this end, a multi-source heterogeneous information fusion prototype network based on compact-sparse representation (CS-MHPN) is proposed in this study for rotating machinery few-shot fault diagnosis. Specifically, with the designed multi-branch prototype network, CS-MHPN can effectively mine and fuse the rich and complementary fault-related information embedded in the multi-source heterogeneous monitoring data to achieve a more comprehensive assessment for equipment health conditions. Additionally, in combination with the proposed compact-sparse composite loss, the intra-prototype compactness and inter-prototype separability can be further constrained, thus improving the diagnostic performance of the model in the few-shot setting. Extensive experiments constructed on the cylindrical roller bearing dataset validate the feasibility and effectiveness of the proposed CS-MHPN.