Auto Fault Detection Based on Swin Transformer and Mask R-CNN

In recent years, deep learning has been widely used in fault detection. However, existing training sets often consist of fixed-size data, leading to poor continuity and accuracy when detecting faults of varying scales simultaneously. Aiming at these problems, this paper proposes a fault detection method based on Swin Transformer and Mask R-CNN. Firstly, we collect multi-scale fault data and apply rotation, noise addition, and other augmentation techniques to construct a diverse training set. Then, Swin Transformer is utilized as the backbone of Mask R-CNN to effectively extract hierarchical features of faults at different scales. Finally, Mask R-CNN performs instance segmentation to ensure the independence of identified faults. Comparative experiments with traditional coherence, curvature, U-Net, and standard Mask R-CNN on New Zealand seismic data show that the proposed method significantly improves fault integrity and continuity.