Abstract: Accurate 3D fault identification remains a key technical challenge in seismic interpretation. Traditional manual interpretation methods are inefficient and highly subjective, while existing deep learning-based approaches suffer from incomplete fault continuity characterization and sample imbalance. To address these issues, we propose a 3D seismic fault intelligent recognition method based on an improved V-Net architecture and adaptive hard example mining. The architecture adopts an improved upsampling strategy that combines convolutional channel transformation with nearest-neighbor interpolation to address the checkerboard artifacts introduced by deconvolution-based upsampling in conventional V-Net. This strategy effectively mitigates the interference of artifacts on fault boundary features, improves the recognition accuracy of small-scale faults and fault boundaries, and simultaneously reduces parameter redundancy and computational cost. To tackle the challenges of sparse fault samples and ambiguous boundaries in seismic data, we design an adaptive hard example mining loss function that incorporates geological prior knowledge. This loss function employs a dynamic threshold mechanism to identify hard examples and applies a differentiated weighting strategy to fault regions, enabling the model to more effectively learn fine-scale fault features and alleviate the imbalance caused by the scarcity of fault samples. Additionally, a continuity constraint term implemented via gradient penalty is introduced to ensure the geological plausibility of the predicted fault structures. Experimental results on both synthetic and field seismic data demonstrate that the proposed method achieves superior accuracy, continuity, and generalization compared with the conventional V-Net, with an F1 score of 0.7962 and a mIoU of 0.7891. This work provides a new technical pathway for automatic 3D seismic fault identification and broadens the application potential of deep learning in geophysical exploration.