Abstract: To address the issues of insufficient fault feature extraction, single-source monitoring signals, and low signal quality in rotating machinery fault diagnosis, an intelligent fault diagnosis method based on multi-source data fusion and an improved convolutional attention module was proposed for accurate fault cause identification. Vibration, current, and torque signals from three different physical properties were collected and processed through filtering and noise reduction to construct a multi-source input vector. The traditional convolutional attention module was structurally enhanced to improve the model’s perception and extraction capability for key features. An end-to-end deep learning fault diagnosis network model was constructed and experimentally validated on the public Paderborn bearing dataset. The results demonstrate that the integration of multi-physical signals during model training significantly enhances diagnostic accuracy. Under varying noise conditions, the diagnostic performance of the improved convolutional attention module is consistently demonstrated to surpass that of the original module. The proposed method achieves a fault diagnosis accuracy of 99.9%, which represents a 0.1% improvement over the existing state-of-the-art results on the same public dataset, thus verifying its effectiveness and advancement. A new technical pathway for rotating machinery fault diagnosis is provided, and valuable reference for intelligent diagnostics in related fields is offered.