Abstract: For the non-stationary vibration signals of gearboxes in wind farm clusters with centralized and remote operation and maintenance scenarios, a time-frequency domain collaborative feature extraction and attention-driven fault diagnosis model is proposed. This model undergoes hierarchical 5-fold validation on long-term online data from a single 3MW unit in a regional wind farm, comprising 1,200 samples across four states. The model achieves an overall accuracy of 98.5%, macro-averaged recall of 97.8%, and macro-averaged F1 score of 98.1, showing improvements of 3.3% and 2.4% in accuracy compared to one-dimensional convolution and bidirectional LSTM, respectively. Results demonstrate that the proposed method significantly enhances recall for tooth breakage and bearing looseness under variable-speed and load disturbance conditions, effectively reduces confusion between wear and normal states, and exhibits practical feasibility for engineering deployment.