Research On Seismic Fault Identification Method Based On Deep Learning

Fault interpretation is the core task of seismic data interpretation,and accurate identification of faults is essential for oil and gas exploration and development.With the increasing demand for oil and gas,many researchers have proposed various effective fault identification methods.Traditional automatic fault identification methods have improved the identification speed compared to manual interpretation methods,but they require manual setting of certain parameters,resulting in lower accuracy of fault identification.With the development of artificial intelligence,using deep learning methods to improve fault identification has become a research direction.Therefore,this thesis focuses on the intelligent identification of faults and specifically carries out the following work:(1)A method for using deep learning models to identify faults is proposed.To address the problem of limited fault feature extraction ability of UNet++ and insufficient continuity of fault identification effect,a CBAM-ASPP-UNet++ fault identification method is proposed.Based on the UNet++ network,a CBAM module is added to enhance noise suppression,improve attention to fault features,and an ASPP module is used to improve the feature extraction ability of multiscale faults.At the same time,synthetic seismic data is used as experimental data to save annotation costs.The joint loss function constructed by introducing cross-entropy loss and Focal Tversky Loss solves the problem of class imbalance.The experimental results on the synthetic seismic dataset show that the segmentation performance of CBAM-ASPP-UNet++ surpasses several classical segmentation models such as FCN,Seg Net,and UNet.Compared with UNet++,the proposed method improves the Precision,MIo U,and F1 indicators by 5.85%,1.09%,and1.83%,respectively.(2)A fault identification method combining transfer learning and semi-supervised learning is proposed.To address the problem of complex and diverse structures of real seismic data,difficult annotation,and limited labeled data,and the weak generalization ability of fault models trained on synthetic seismic data.This thesis uses synthetic seismic data for transfer learning to enhance the model’s generalization ability.Semi-supervised learning is used to solve the problem of deep learning methods relying heavily on data quantity,and to improve the model’s segmentation performance.A fault optimization postprocessing method is proposed to enhance the continuity of faults,make the predicted results closer to the real labels,and reduce the interference of pseudo-label errors on network learning.The experimental results on the real seismic dataset show that semisupervised transfer learning effectively solves the problem of difficult manual annotation,and the method has good performance and generalization ability.(3)A deep learning-based seismic fault interpretation system is designed and developed.The fault interpretation model and methods proposed in this thesis,as well as classical convolutional neural networks,are integrated into the system,and the reading and multidimensional visualization display of seismic data,model training,fault prediction,and fault post-processing functions are implemented using Py Qt5 and Python.At the same time,application research is conducted in this field,which is in line with actual production conditions.