Abstract: Deep learning is a powerful tool for fault identification based on seismic data. However, traditional methods are plagued by poor dataset quality, excessive resource consumption, and lengthy training cycles. To address these challenges, we propose a 3D fault identification method integrating an improved U-Net and knowledge distillation. An enhanced U-Net model, working as the teacher model, is integrated with the atrous spatial pyramid pooling (ASPP) structure to construct a lightweight student model. The model is then optimized through knowledge distillation. By adjusting network training hyperparameters and knowledge distillation loss parameters, the model acquires richer fault information and thereby enhances its network performance. Through transferring knowledge from the complex teacher model to the lightweight student model, this approach significantly reduces computational complexity while maintaining high recognition accuracy. Synthetic and field data tests demonstrate that the knowledge-distilled student model outperforms both the undistilled student model and the independently trained teacher model in terms of recognition accuracy and fault continuity, fully verifying the feasibility and effectiveness of this method.