Abstract: Small faults in coal-fields are major triggers of geohazards, including coal-gas outbursts and mine water inrushes, and thus require accurate detection and fine-scale characterization. Conventional structural interpretation and seismic-attribute-based methods (e.g., coherence, variance, and curvature) often suffer from missed and false detections under low signal-to-noise ratios and complex geological settings. Recent advances in machine learning and deep learning — spanning SVM and random forests to CNN/RNN models and U-Net variants — have enabled increasingly automated small-fault identification from 3D seismic data. The evolution from attribute-driven interpretation to data-driven classification and end-to-end detection is summarized, representative studies are highlighted, and two key challenges are discussed: limited labeled samples and weak fault signatures masked by strong coal-seam reflections. Future directions are outlined, including physics-informed forward modeling, data augmentation and transfer learning, multi-attribute fusion, and geology-constrained interpretation.