In the context of wind turbine gearbox fault early warning, this paper proposes a method based on graph attention and temporal convolutional networks to address the issue of insufficient data information mining. By establishing physical connections for each feature point in both temporal and spatial scales, the method expands the feature dimension to enhance fault warning accuracy. Graph attention network captures spatial relationships, while temporal convolutional network improves temporal feature capturing. Experimental results using real data from a wind farm show that the proposed method can issue fault warnings 122 hours in advance, outperforming other methods by 52 to 63 hours with reduced prediction errors (1. 05% to 3. 76%). The approach also enhances result interpretability using t-SNE and probability density curve analysis.