Abstract: In machine learning-based fault identification using seismic data, complex structural interference and strong noise may easily lead to missed detection of small faults or misclassification of noise as faults. To address this problem, a fault identification model guided by along-layer seismic attributes, termed discrete wavelet transform−singular value decomposition−U-Net (DSU), was proposed with the Yunnan mining area as the study area. The key features of the model are as follows. First, to tackle the problem of noise interference and weak response of small faults, three mother wavelets with orthogonality, compact support, and high vanishing moments—Daubechies 4 (Db4), Symlet 4 (Sym4), and Coiflet 3 (Coif3)—were selected as candidates. Fast Fourier Transform (FFT) and Discrete Cosine Transform (DCT) were employed to compute the energy differences between fault and non-fault signals in the frequency and spatial domains, respectively, yielding two quantitative metrics for mother wavelet selection: the frequency-domain energy difference index CFFT and the spatial-domain energy difference index CDCT. The comparison showed that the Db4 mother wavelet achieved a CFFT value of 0.0358, a CDCT value of 2.4635, and a 35% improvement in the signal-to-noise ratio at fault locations over the original data in the Yunnan mining area, all of which were superior to those of Sym4 and Coif3. Therefore, Db4 was selected as the optimal mother wavelet for Discrete Wavelet Transform (DWT). While suppressing strong noise, seismic attributes were further extracted along the target layer using a 5 × 5 sliding window, effectively reducing the interference of non-target-layer noise on fault signals. Second, to address the significant correlation and information redundancy among the 20 along-layer seismic attributes in the training dataset, Singular Value Decomposition (SVD) was employed for dimensionality reduction. The optimal decomposition dimension was determined using the hard threshold method based on the criterion that the cumulative energy contribution rate should be no less than 90%, thereby reducing the attribute dimension from 20 to 6. Six core attributes were selected: original amplitude, root mean square amplitude, chaos volume, dip deviation, variance, and instantaneous phase. This approach effectively mitigated the risk of model overfitting caused by highly correlated attributes while retaining the majority of feature information. In U-Net classification network, to address the class imbalance problem arising from the substantial disparity between fault and non-fault pixel counts, a weighted cross-entropy loss function was introduced, in which the class weights were set inversely proportional to the sample sizes of the two classes, thereby imposing a higher misclassification penalty on fault samples. A training dataset was constructed from Area A, which is dominated by large-throw faults, and a testing dataset from Area B, which is dominated by small- and medium-throw faults. Three comparative experiments were designed using U-Net, DWT−U-Net, and DSU models. The results show that the DSU model achieves an accuracy of 0.93131, a precision of 0.95341, a recall of 0.90695, and an F1 score of 0.9296, all of which represent significant improvements over the U-Net-only baseline. The model maintains effective identification capability for small faults with throws of less than 5 m in Area B; the small fault F67 was confirmed by roadway exposure data to be consistent with the predicted result. Furthermore, the model output was converted into posterior probability values, enabling a quantitative reliability assessment of the fault identification results and providing interpreters with an intuitive reference map of fault development probability.