Abstract: Seismic fault identification is critical for hydrocarbon exploration and development. However, existing fault identification methods still encounter several challenges when processing complex seismic data, including inadequate multi-scale feature coverage, insufficient modeling of long-range fault continuity, and susceptibility to noise interference. To address these limitations, a 3D seismic fault identification model based on dual-attention multi-scale fusion U-Net (DAMFU-Net) was proposed. Firstly, the multi-scale feature fusion module served as the basic feature extractor. Through a multi-branch structure, it initially captured the features of faults with different scales in the seismic data, providing rich multi-scale feature representations for subsequent processing. Furthermore, a dilated fusion module further expanded the receptive field on this basis. Through multi-expansion rate convolution, it explored long-range spatial dependencies and effectively modelled the global continuity characteristics of the fault structure. Finally, based on the multi-scale and long-range features extracted in the first two modules, the dual-attention parallel mechanism achieved further feature selection and fusion in the skip connections. Through the collaborative attention calculation in the channel and spatial dimensions, it enhanced the model’s sensitivity to fault features and its ability to resist noise interference. The experimental results show that, compared with the four currently mainstream methods for fault identification, the DAMFU-Net model has an accuracy improvement of 1.66% over the MARU-Net model in identifying fractures. It also demonstrates excellent noise resistance performance. Experimental verification on the F3 and Kerry-3D datasets has shown that this method has good generalization performance.