To address the challenge of diagnosing faults in motor rolling bearings under low-speed and heavy-load conditions where strong noise interference is present, this study proposes an intelligent diagnosis method that integrates vibration, temperature, and sound as multi-modal signals. A diagnosis model is constructed based on a GRU-Transformer cross-modal fusion framework: gated recurrent units (GRU) are used to extract temporal features of each modality, and Transformer multi-head attention is employed to achieve deep cross-modal fusion and global dependency modeling. Industrial background noise from real-world settings is introduced into the experiments to create interference scenarios. The results demonstrate that the model achieves excellent performance in identifying various fault types; specifically, the F1-score reaches 0.96 under noise-free conditions and remains at 0.95 even under strong noise, indicating minimal performance degradation and outstanding anti-interference capability, with promising application prospects in engineering.