Fault detection based on U-Net and GNN integration

Abstract

Current fault detection methods mainly take advantage of a convolutional neural network, simplified U-Net, for seismic image semantic segmentation, which is a computer vision task aimed at generating a dense pixel-wise segmentation map of an image, where each pixel is assigned to a specific class or object. However, these methods face challenges such as unclear segmentation boundaries and limited receptive fields, which hinder the model's capability to detect all features of the fault. To address these issues, this paper proposes a method based on the integration of Graph Neural Network (GNN) and U-Net, referred to as GNU-Net, which incorporates fault graph structures and graph computation modules into U-Net. This approach aims to expand the receptive field, clarify segmentation boundaries, and optimize fault segmentation and continuity analysis through a joint loss function. The method consists of five modules: graph construction, data augmentation, fault segmentation, fault continuity analysis, and joint training. The graph construction module grids the fault labels with a grid edge length of 8 voxels, sets vertices at the locations where the fault intersects the grid, and selects representatives from the voxels labeled as faults within each grid as vertices. The dilation function is applied to expand the fault voxels in the labels and construct a velocity field, in which the geodesic distance between vertices is computed to establish connections between vertices. Data augmentation of the 3D seismic data and fault labels is achieved by rotating the images. A custom rotation matrix is applied to transform the vertex feature coordinates of the graph structure, significantly increasing the size and diversity of the training dataset. The fault segmentation module employs a simplified U-Net with square convolution kernels, and the segmentation loss includes both Dice and binary cross-entropy losses. The fault continuity analysis module leverages a graph neural network with irregular convolution kernels to capture macroscopic fault features and enhance fault continuity, employing a connection loss based on binary cross-entropy. Compared to the original U-Net, the modified GNU-Net achieves a fault detection accuracy of 97.39 % on the testing set, an improvement of 3.96 % over the original U-Net.