3D Seismic Fault Recognition Based On Convolutional Neural Network

With the rapid development of the economy,the demand for oil and gas energy is on the rise.The identification of faults as hydrocarbon migration channels has long been the focus of seismic exploration.The method of fault identification has evolved from traditional manual marking to automatic attribute marking and presently to intelligent deep learning interpretation.The advancement in technology has resulted in higher accuracy and efficiency in fault identification and has also brought new research ideas and methods for earthquake prediction and geological exploration.With the evolution of computer vision technology,an increasing number of scholars have applied deep learning algorithms in the field of artificial intelligence for automatic fault recognition to further improve the effectiveness of fault recognition.This thesis is based on the research idea of semantic segmentation,which employs deep learning technology to automatically and intelligently identify faults.The convolutional neural network can recognize earthquake faults end-to-end.However,there remain issues such as a small proportion of fault information in seismic data,the absence of uniform fault scale and scattered fault shape.On the one hand,the unbalanced proportion of positive and negative samples can be mitigated by improving the loss function.On the other hand,multi-scale features can be integrated and the feature extraction ability can be enhanced to comprehensively learn fault information.This thesis utilizes two different UNet-based convolutional neural network models to identify seismic fault data in two and three dimensions.The specific work carried out is as follows:(1)Two-dimensional fault recognition method based on improved UNetAiming at the problem of irregular earthquake size and small fault characteristics are not obvious,this thesis proposes a two-dimensional fault identification method based on improved UNet.Specifically,a CSPNet network model is developed based on the UNet architecture,which can adopt the multi-scale fusion strategy for the size irregularity of seismic data.In order to improve the accuracy of fault identification,an unbalanced loss function is introduced to optimize the unbalanced proportion of fault and non-fault samples in seismic data.In view of the standard UNet single jump connection structure,irrelevant information is introduced and the high level global context information and low level high-resolution local information are not effectively integrated.The global pyramid guide module is also added through jump connection between the encoder and decoder to provide context information of different levels and enhance the expression ability of seismic fault characteristics of the network model.Finally,the results of IOU(72.5%)and F1(84.0%)in synthetic seismic data are obtained by the algorithm,which verifies the effectiveness and reliability of the algorithm in the field of fault identification.(2)Attention mechanism and 3D-UNet fusion 3D fault recognition methodThe Res UNet-MAM model is proposed for the identification and segmentation of seismic faults using the attention mechanism.This model is based on the 3D-UNet and incorporates the 3D attention mechanism and residual convolutional module to enhance the feature expression ability of the model and make it pay more attention to fault features.Finally,the proposed model is compared with the classical semantic segmentation model and the existing attention mechanism,which proves the validity and reliability of the proposed method.(3)Application of Res UNet-MAM model to actual seismic dataFinally,this thesis applies the proposed model Res UNet-MAM to the actual seismic data of F3,and compares it with the traditional method and the existing fault Seg.Through the identification results of the actual seismic data,the superiority and practical value of the proposed method are verified in terms of the strike,continuity and clarity of the fault.This thesis contains 43 pictures,12 tables and 88 references.