Physics-model-guided frequency spectrum dynamic aggregation network for gearbox signal denoising under multi-source disturbances.

As a critical component of rotating machinery, the stable operation of gearboxes is essential to industrial systems. In practice, complex operating conditions cause fault-related features to be heavily obscured by noise, posing significant challenges for fault diagnosis. Most denoising tasks assume that experimental signals are noise-free and use Gaussian white noise to simulate noise interference. However, this overlooks the fact that acquired signals are already affected by various coupled disturbances and inherently contain a certain level of noise. To address this issue, a physics-model-guided denoising method for gearbox signals is proposed. First, a dynamic gear transmission model is established to generate noise-free reference signals under ideal conditions, and its key parameters are optimized using the grey wolf optimization algorithm to reduce discrepancies with the real signal. Subsequently, a novel deep denoising network is developed, in which a frequency spectrum dynamic aggregation mechanism adaptively aggregates spectral features to emphasize informative frequency components while suppressing interference from spurious frequencies. Additionally, an adaptive hybrid loss function integrating time-domain, frequency-domain, and phase information is designed to enhance robustness against diverse noise types. The proposed method is validated on a gearbox dataset. Experimental results show that it outperforms other models on most evaluation metrics. Even under severe noise-coupling conditions, the average RMSE of the denoised signals is maintained at 0.0498. In fault diagnosis experiments, the proposed method achieves an average accuracy of 97.65%, which is approximately 3.5% higher than that of other U-Net-based denoising approaches.