Research On Seismic Fault Identification And Profile Prediction Based On Deep Learning

In recent years,the rapid development of neural networks has promoted the wide application of deep learning in society.Now deep learning has been used in social production,promoting the development of social productivity.With the development of oil exploration technology,there are now higher requirements for fault identification.Over the past few years,fault identifications are many traditional algorithms,such as ant tracking,coherent algorithm,seismic curvature fault detection technology,etc.However,in recent years,the research direction has gradually shifted to the direction of deep learning.The use of deep learning technology for fault identification has become a research trend of seismic data processing in the future.The complex topography of China makes it difficult to collect seismic data,and it is difficult for researchers to carry out exploration work smoothly under the complex terrain.In this paper,a spatiotemporal prediction model is designed.According to the limited resources and data,the seismic data of the unexplored areas are predicted.After obtaining the predicted seismic data,fault identification is carried out.This paper mainly uses machine learning technology to study seismic data prediction and fault identification.The research content is as follows:For earthquake data prediction problems,earthquake data prediction is also a kind of spatiotemporal prediction,so we established a spatiotemporal prediction model to predict earthquake data.With the development of deep learning,spatiotemporal prediction learning has many applications,such as fault recognition,weather prediction,and human pose prediction.Describe the continuous data problem as a spatiotemporal series prediction problem where both the input and the predicted target are spatiotemporal series.In multi-step spatiotemporal predictions,the image also becomes blurry as the step size increases.Based on the above discussion,we designed the SIM block and the LIM block,which interacts with the current input with the hidden state,to update the current input and hidden state,and capture the important parts of the instantaneous state to overcome this difficulty.LIM blocks interact with memory cells to update memory cells and hide states,capturing deep semantic information through attention blocks and multi-scale feature extraction.SIM blocks and LIM blocks are added to Conv LSTM,and an IM-LSTM network framework is proposed,and an end-to-end trainable model for prediction problems is established with it.Experiments demonstrate the effectiveness and flexibility of the model.Experiments on two datasets have proved that the model can predict a clear future with some generalization.It is superior to other advanced methods in the indicator LPIPS.For the fault identification problem,we have improved it on the classic network U-net.A multi-branch parallel structure M-block(Multibranch block)is redesigned for the decoder’s convolutional block.It captures multi-scale contextual information,and the multi-branched parallel structure delivers high-performance benefits.In addition,self-attention blocks and attention gating mechanisms have been added to the decoder section,which can compensate for the shortcomings of CNN network locality.Self-Attention enables the attention module to flexibly focus on different areas of the image by performing a weighted averaging operation of the input feature context.At the same time,it can overcome the limitations of the locality of convolutional networks and bring more possibilities to convolutional neural networks.The improved network combines the advantages of weight sharing in traditional convolution with the advantages of self-attention dynamic calculation of attention weights,fully extracts the features of images,and improves the accuracy of fault recognition in experiments.Both of the above problems have been experimentally tested to test the conjecture.After comparison,it is concluded that the IM-LSTM network proposed in this paper based on the seismic data prediction problem performs well on the Moving MNIST dataset and the 3D seismic dataset.The improved U-Net network based on the fault recognition problem outperforms the general model on the test dataset of predicted frames and Wu and performs well.