<p>Precise fault boundary delineation is crucial for oil and gas exploration and development. Meanwhile, artificial intelligence algorithms have achieved remarkable results in fault detection tasks. In particular, the application of attention mechanisms in deep learning has significantly improved feature extraction in seismic interpretation. However, relying on a single type of attention module may cause the network to concentrate too narrowly on certain training features, which can lead to overfitting and hinder generalization. Although integrating multiple models can improve overall performance and robustness, it incurs a relatively high computational cost. In response to the above problems, we propose a novel two-branch network architecture that adds a separate encoding and decoding path and introduces a new lightweight attention mechanism on the basis of the existing U-Net structure. Two separate encoding and decoding branches can be flexibly learned to obtain different fault characteristics during training, thereby reducing implicit prediction errors and improving the generalization ability and stability of the model. This study used the FaultSeg3D synthetic dataset for training, with input features being three-dimensional seismic amplitude volumes and output being corresponding fault probability annotations. Validation was then performed on two publicly available field seismic datasets: the Netherlands offshore F3 block and the Kerry-3D dataset. The experimental results show that this mechanism enhances the representation of fault continuity. The two-branch network architecture focuses on a variety of typical fault characteristics. In particular, in the initial training iterations, it accelerates the convergence of multiple evaluation metrics compared to each branch used independently. After applying the complementary advantages of the two network branches, the combined network more effectively identifies key faults in real seismic data and enhances the representation of detailed fault features.</p>