Earthquake fault detection is a key task in establishing a georeservoir model, and high-precision, high-efficiency and high-resolution detection results are an important guarantee for accurate geological structure interpretation. With the development of artificial intelligence algorithms in the field of object detection, using the deep convolutional neural network to predict fault locations in seismic data has become a research hotspot. neural network establishes a nonlinear mapping relationship from seismic data to target faults by extracting potential features of faults from seismic reflection information and suppressing interference from background noise, thereby achieving end-to-end fast prediction. The current mainstream intelligent detection method for tomography is to train images to segment networks (such as U-Net and its variants) using large-scale synthetic data. However, large-scale network training and high-parameter neural networks rely on hardware devices with high computing performance, which is expensive to calculate and poorly portable, which affects technological development and industrial promotion. In addition, the synthetic data is a forward simulation of artificial models. The simplified model of geological conditions is difficult to reflect the complex actual structure of the underground. Therefore, exploring the use of small sample real seismic data to train lightweight neural networks to achieve high-resolution fault recognition is a very meaningful research direction.
Dr. Wang Tianqi and supervisor of the Key Laboratory of the Institute of Oil and Gas Resources, Institute of Geology and Geophysics, Chinese Academy of Sciences, proposed a high-resolution seismic fault interpretation network (FaultAdvNet) based on the combination of adversarial game mechanism and regularization technology. On the one hand, the network uses the global feature fusion module to comprehensively consider the target and background information, and on the other hand, it uses the data-driven auxiliary network as an implicit regularization constraint term. It is necessary to use only a limited training sample to achieve high-resolution target geological detection.
FaultAdvNet's network architecture consists of three parts (Figure 1): a segmentation module that performs fault recognition based on seismic inputs, a feature fusion module that integrates target features (fault) and background information (stratigraphic reflection) into a global feature map, and a discriminator module that distinguishes fault prediction results from fault labels. Among them, the segmentation module (Fig. 1a) is a lightweight prediction network with a parameter volume of only 1.5% of the conventional U-Net model. The feature fusion module (Fig. 1b) is a feature enhancement process designed based on the geophysical concept of faults. By highlighting the main stratigraphic interface and fault staggered boundaries, the network is urged to pay attention to the reflection characteristics of faults and surrounding sediments during the training process, thereby overcoming common fault segmentation problems (such as blurred boundaries, poor continuity, and misjudgment of the inclined formation as faults, etc.). The segmentation module (Fig. 1a) and the discriminator module (Fig. 1c) improve each other's performance through adversarial game training. Specifically, the former deceives the latter by generating prediction results closer to the label, and the latter punishes the former by discriminating the difference between the prediction results and the label. Quantifying the penalty value can constitute a regularized loss function, which is the weighted pixel-level prediction error and image-level prediction error, reflecting FaultAdvNet's overall control over local target information and global macro information.
Figure 1 Network architecture of the FaultAdvNet model. (a) segmentation module; (b) feature fusion module; (c) discriminator module
through module function comparison analysis (Figure 2), they evaluated the importance of each module in the FaultAdvNet model and their synergistic relationship, verifying the superiority of this network architecture. Through comparison with the U-Net model, it is argued that neural network architectures designed for specific tasks are more suitable for solving professional problems.
Figure 2 Analysis of module function of FaultAdvNet model during training. (a) FaultAdvNet model for initial training of segmentation modules; (b) FaultAdvNet model; (c) FaultAdvNet model for featureless fusion modules and discriminator modules; (d) U-Net model
Gulf of Mexico case study shows that compared with traditional methods (such as coherent algorithms) and other intelligent methods (including multiple CNN deep learning methods), FaultAdvNet's three-dimensional prediction results have good continuity, clear boundaries and higher confidence (Figure 3).As an interdisciplinary study of computer vision and geophysics disciplines, this method has broad application prospects in various geological recognition tasks (such as river channels, salt hills, gas chimneys, etc.).
Figure 3 A three-dimensional view of the fault prediction results. (a) Three-dimensional seismic data in the Gulf of Mexico region; (b) prediction results of third-generation coherent algorithms; (c) prediction results of three-dimensional convolutional neural networks; (d) prediction results of U-Net model; (e) prediction results of FaultAdvNet model
research results were published in the international academic journal Geophysics (Wang T Q and Wang Y F*. High-resolution seismic faults interpretation based on adversarial neural networks with a regularization technique[J]. Geophysics, 2022, 87 (6): IM207–IM219. DOI:10.1190/geo2021-0383.1). The research is funded by the National Natural Science Foundation of China (12171455), the from 0 to 1 (ZDBS-LY-DQC003), and the key deployment project of the Institute of Geology and Geophysics, Chinese Academy of Sciences (IGGCAS-2019031).
Editor: Fu Shixu
Proofreading: Wanpeng
