Research On Seismic Exploration Signal Classification Method Based On Deep Learning

Seismic prospecting is a very important tool for hydrocarbon exploration,and the processing,interpretation and analysis of seismic signals can help to understand the reservoir characteristics and predict the distribution of oil-gas reservoir.The seismic signal classification method combines pattern recognition methods and seismic response characteristics to map the seismic signals into different categories,which is convenient to understand the spatial distribution of reservoirs intuitively.However,the classification results are largely dependent on the selection of features.The large number of redundant stratigraphic response features often make it difficult for interpreters to select useful features effectively,resulting in high interpretation workload,low efficiency and high cost,and there is a certain degree of human subjectivity.Deep learning algorithms can automatically extract implied features from a large number of seismic signals,and have become a popular research topic in the field of seismic signal classification.In this thesis,focusing on the problem of insufficient mining if stratigraphic reflection information and lack of interpretation labels,the research on the classification method of seismic signals based on deep learning is carried out.The main work is as follows:(1)To address the problem of inadequate mining of stratigraphic reflection information,this thesis proposes an unsupervised graph embedding classification method based on the information-rich pre-stack seismic signals to study the characteristics of stratigraphic response with azimuthal changes.This method constructs a graph structure to describe the wide azimuth pre-stack data,so that the temporal attributes of the seismic signal in the vertical direction and the azimuth information in the horizontal direction can be used to automatically extract the corresponding features of the stratum using the graph embedding algorithm,and further,the graph embedding features are unsupervisedly clustered using K-means to obtain the final classification results.The experimental results on the real seismic data verify the superiority of the algorithm in the problem of classification of pre-stack seismic signals.(2)To address the problem of difficult feature extraction of post-stack seismic exploration signals,this thesis proposes a deep clustering method incorporating temporal information.CNN and Bi LSTM are fused in the feature extraction stage.CNN is used to extract rich local information,and Bi LSTM is used to extract the temporal dependence of seismic signals.Aiming at the problem of few seismic data interpretation labels,a selfsupervised learning method is used to mine seismic signal features.That is,using the clustering result as a pseudo-label to pass to a classifier,the classifier also accepts the feature vector passed by the feature extraction module,conducts supervised training,and iteratively updates the classification parameters and feature extraction parameters,so that the features extracted by the network are more cluster-friendly while maintaining the information content.By applying the algorithm to the physical model data,it is proved that the algorithm can obtain better classification results for post-stack signals.