To enhance model interpretability in aerospace bearing fault diagnosis, this paper proposes a Physics-Guided Multi-Scale Feature Fusion Network (PGMSFFN). By introducing continuous wavelet transform in the initial layer, the model achieves multi-scale time–frequency decomposition with clear physical meaning. A dynamic weighting mechanism based on energy distribution further enhances fault feature representation. Experiments on the HIT aerospace bearing dataset show that PGMSFFN surpasses existing Wavelet Kernel Network and Wavelet Kernel Attention Network models in both accuracy and stability. Feature visualization confirms that the method effectively captures physically relevant fault features, providing a more interpretable foundation for practical diagnosis.