A Lightweight Multi-Scale Attention Gated Network for Wind Turbine Bearing Fault Diagnosis Under Complex Operating Conditions

Under strong noise interference and non-stationary operating conditions, vibration signals of wind turbine bearings exhibit complex characteristics, making fault diagnosis difficult. Meanwhile, practical fault diagnosis applications require models that balance diagnostic accuracy and computational efficiency. Existing deep learning methods often suffer from high computational cost, weak noise robustness, and limited generalization in few-shot scenarios. To address these issues, this paper proposes a lightweight fault diagnosis model named MDSAG. The model employs multi-scale depthwise separable convolutions to extract features across different frequency bands while reducing computational cost. Channel attention enhances informative feature responses and suppresses noise interference, while coordinate gating captures temporal variations to improve feature representation under non-stationary conditions. These components are integrated into a unified lightweight framework for efficient feature extraction and robust representation learning. Experiments on the Case Western Reserve University dataset and a real wind turbine bearing dataset demonstrate that MDSAG achieves high diagnostic accuracy under varying noise levels and few-shot settings while maintaining low computational complexity, highlighting its potential for practical industrial fault diagnosis.