<p>Accurate identification of seismic faults is an important step in the development process of underground oil and gas reservoirs. Although semantic segmentation technology has been widely applied in the field of fault identification, this method still has problems such as blurred fault features, discontinuous identification of fault lines, and misidentification and missed identification. In response to these problems, this paper proposes an intelligent fault identification method based on multi-scale feature fusion and a dual-decoder architecture. Based on the U-shaped encoder-decoder architecture, the method constructs a dual-decoder network. The feature fusion of the dual-decoder results ensures comprehensive fault information reconstruction and continuous fault identification. Meanwhile, a three-channel fusion feature module is designed to extract the pooling features, convolution features and residual features of the input seismic data, highlighting the target fault line and enhancing the model’s identification ability on complex faults. Develop a four-scale kernel path module to replace the traditional skip connection operation, utilizing multi-kernel paths to adapt to hierarchical fault features, and enhance the model’s extraction ability of edge-hidden faults and long-distance faults. The accuracy of this method in fault identification tasks on synthetic datasets reaches 95.03%, which is 1.8% higher than the U-Net network with a single encoder-decoder structure. The experimental results on the F3 dataset in the Netherlands show that this method provides clearer and more accurate characterization of fault lines in real seismic data. It effectively reduces false identifications and missed detection of faults, while improving the model’s performance in fault localization and segmentation capabilities across multi-scale fault structures.</p>