Bearing fault diagnosis is crucial for ensuring the operational reliability and safety of rotating machinery. However, traditional diagnostic methods face significant challenges: strong background noise often obscures weak fault signals, and reliance on single-sensor information provides an incomplete picture of complex fault states. These limitations can lead to decreased diagnostic accuracy and reliability. To address these issues,  this study proposes a novel convolutional sparse bearing fault pattern recognition method based on multi-source heterogeneous information from vibration and sound signals. Firstly, embedding a convolutional sparse representation layer within the deep learning framework, which adaptively extracts key quasi-periodic fault impulses from raw multi-source signals and suppresses noise and irrelevant background features through an end-to-end learned convolutional dictionary and sparse constraints; secondly, designing a novel fusion convolutional layer that integrates standard convolution and dilated convolution to capture local fine structures and broader contextual and multi-scale information in parallel, thereby achieving deep and effective fusion of purified feature streams from different sensors.  Finally, a classification stage utilizes these comprehensively fused features to perform the bearing fault diagnosis. Experimental results on multi-source heterogeneous datasets from two bearings types demonstrate that the proposed M-DSRN method achieves outstanding fault recognition performance, compared with traditional rivals including deep neural networks and Sparse Representation Classification. This fully validates its effectiveness and advancement in achieving high-precision fault diagnosis under conditions involving strong background noise and multiple fault types through enhanced feature extraction and multi-source information fusion.

Bearing Fault Diagnosis,Deep Learning,Convolutional Sparse Representation,Multi-source Information Fusion,Pattern Recognition