Research On Analysis Method Of Prestack Seismic Reflection Pattern Based On End-to-End Deep Clustering

Prestack seismic reflection analysis is a method to determine the subsurface rock structure and its characteristics by analyzing the reflection and refraction phenomena generated when seismic waves propagate in the subsurface rock and finally generating a pre-stack seismic section.Ultimately,this method generates prestack seismic phase diagrams.By analyzing these diagrams,geological interpreters can intuitively understand the distribution of rock types and identify the location of oil and gas.Therefore,prestack seismic reflection pattern analysis provides important support for the development of geological interpretation and oil and gas exploration.Prestack seismic data is vast and contains more geological information,so it is usually subjected to feature extraction before clustering identification.However,the isolated optimization of feature extraction and clustering may lead to suboptimal results,making it challenging to simultaneously accomplish feature extraction and clustering in seismic reflection pattern analysis.Deep clustering is a novel clustering framework that integrates feature learning and clustering into a unified framework,capable of directly clustering the original seismic images.Therefore,this thesis aims to investigate end-toend deep clustering algorithms for prestack seismic data and proposes prestack seismic reflection pattern end-to-end deep clustering methods based on log-normal mixture variational autoencoder and focal loss convolutional autoencoder,with the following innovative research results.(1)To address the issue of asymmetric distribution of deep features in seismic data,which traditional deep clustering models cannot accurately represent,and to overcome the problem of difficult inference of deep generative models in complex latent structures,this thesis proposes a end-to-end deep clustering model based on the log-normal mixture variational autoencoder.The model replaces the traditional Gaussian mixture distribution with a log-normal mixture distribution in the probability distribution of the feature space,which more accurately models the asymmetric distribution of seismic data in the deep feature space.By inferring which mode of the latent distribution the data point is generated from in the hidden layer,the model infers the category of the data point to achieve deep clustering.In addition,a reparameterization trick-based inference model is constructed,which simplifies the inference model by directly optimizing it.(2)To address the issue that existing deep clustering methods often ignore the representational power of latent features and suffer from the misclassification of samples in the clustering layer,this thesis proposes a joint optimization framework,end-to-end a deep clustering model based on the focal loss convolutional autoencoder.In the reconstruction term,the model uses a weighted sum of mean squared error and binary cross-entropy as the loss function,where the binary cross-entropy loss is mainly used to penalize the structural differences between the original data and the reconstructed data,to enhance the representational power of the latent features.In the clustering term,the model constructs a focal loss to improve the label assignment mechanism,and utilizes self-training in the network to view the target distribution in the clustering layer as the true label distribution,and thus applies the focal loss in an unsupervised manner to further improve the clustering performance.The two deep clustering algorithms proposed in this thesis were verified on prestack seismic data from both synthetic models and actual work areas,and the results demonstrated that our method can provide more and accurate geological information.