As a core component of motor equipment, bearings are susceptible to faults such as inner ring wear, outer ring spalling, rolling element damage, and cage fracture under complex industrial operating conditions. Existing diagnosis methods suffer from poor noise adaptability, architectural mismatch with one-dimensional vibration signals, and insufficient integration of expert prior knowledge. This paper proposes a motor bearing fault diagnosis method based on composite denoising and Frequency Domain Separation Convolutional Network (FDSCN). First, a divide-and-conquer composite denoising strategy employing SVD, VMD, wavelet packet adaptive thresholding, and EEMD in parallel accommodates diverse noise scenarios. Second, a dual-stage multi-domain feature fusion strategy performs signal-level fusion of denoised signals and achieves cross-domain information complementation between vibration and working condition features via an attention mechanism. Third, FDSCN is constructed with kernel frequency decoupling, element-wise frequency response fine-tuning, and spatial dynamic filtering, specifically tailored for one-dimensional vibration signals. Experimental results demonstrate that the proposed method achieves an overall accuracy of 98.9% and a macro-averaged F1 score of 98.3%, outperforming various existing methods while maintaining low parameter count and fast inference speed suitable for industrial online diagnosis.

motor bearing; fault diagnosis; deep learning