Application of improved Swin-Transformer model in fault identification

This study proposes an improved Swin-Transformer - based method for seismic fault identification. To address the limitations of traditional approaches and supervised deep learning models, we integrate self-supervised pre-training method with a 2D Swin-Transformer architecture. Seismic slices are divided into patches, with random regions masked, and the model is trained to predict masked parts through unsupervised training. The fault recognition network incorporates dual fully convolutional network (FCN) segmentation heads, coupled with binary cross-entropy (BCE) loss training, to generate fault probability predictions. Experimental results on field seismic data demonstrate superior accuracy in fault localization and morphology identification compared to conventional Unet-based methods. The approach reduces computational costs and offers a practical solution for resource exploration and geological hazard assessment.