<p>Fault diagnosis of bogie traction motors is essential to ensure the operational safety of rail trains. Conventional data-driven models rely primarily on statistical signal correlations and often overlook the domain-specific knowledge regarding mechanical, electrical, and dynamic couplings. This study proposes a multimodal fault diagnosis framework that integrates vibration, current, and rotational speed signals. The raw one-dimensional signals are transformed into spectral correlation (SC) maps, which leverage cyclostationary theory to characterize the modulation components caused by mechanical–electromagnetic interaction. A redesigned ConvNeXt backbone with cross-resolution gated attention (CRGA) is employed to capture global dependencies across frequency resolutions. To guide the feature learning process with physical priors, two physics-inspired regularizers are introduced during training: an energy consistency loss motivated by balanced energy distribution among modalities, and a dynamic correlation loss reflecting the inherent mechanical–electrical coupling. Experiments on the BJTU-RAO and HUST multimodal motor datasets demonstrate that the proposed model achieves superior diagnostic accuracy and robustness.</p>