Abstract: Fault prediction is critical for safe coal mine production. Traditional machine learning methods suffer from poor prediction accuracy when fault features are subtle. This study therefore proposes a WT-U-Net model by combining wavelet transform (WT) with U-Net to improve interpretation accuracy. Fault-related attributes were extracted from post-stack seismic data, where four attributes with low mutual dependence were identified through correlation analysis. The decomposition-reconstruction errors and energy differences of different wavelet basis functions applied to seismic data were then compared. The coif3 mother wavelet was selected for fault detection as its wavelet transform amplified fault-related signatures. The U-Net model was constructed to predict faults in the study area. Results demonstrate that the WT-U-Net model showed higher prediction accuracy than UNet alone on real datasets, with outputs more consistent with manual interpretations and improved convergence. The model also exhibited robustness and generalization in blind tests across other regions. The application of wavelet transform in seismic data denoising enhances fault-related signals, thereby increasing the model's accuracy. This study offers a new solution for intelligent fault identification in coal mining applications.