Accurate fault prediction in coalfield seismic data is important for geological interpretation and the safe and efficient exploitation of coal resources. However, conventional fault interpretation methods and shallow machine-learning approaches usually rely on manually extracted seismic attributes. They often show limited robustness to noise and insufficient capability in characterizing fault continuity, boundary features, and small faults in structurally complex areas. To overcome these limitations, a fault prediction method based on the singular value decomposition–residual convolutional block attention module–U-Net (SVD–ResCBAM–U-Net) framework is proposed. First, SVD was used to denoise the seismic data and improve its quality. Then, residual blocks and a convolutional block attention module were incorporated into the U-Net architecture to enhance fault-related feature extraction and improve prediction performance. Experimental results show that the proposed SVD–ResCBAM–U-Net achieved the best performance among all compared models, with a global accuracy of 0.9556, a mean intersection over union of 0.6241, and a mean boundary F1-score of 0.6796. These results clearly demonstrate the proposed method’s advantages in fault continuity, boundary delineation, and small-fault prediction, underscoring its effectiveness for fault prediction in coalfield seismic data under complex geological conditions.