Characterization Of Shale Gas Layer Reflection Seismic Signals Based On Sparse Autoencoder

Natural gas is a strategical-clean-energy with short supplement in China.To increase the exploration of shale gas is an important part of the national energy strategy.China’s shale gas is widely distributed,deeply buried,early formed,highly evolved and highly heterogeneous.Conventional gas detection methods(such as bright spot technology,AVO,etc.)have limitations.Deep learning has the ability to automatically learn the essential characteristics of data through multiple nonlinear transformations.This paper combines deep learning with seismic signal time-frequency analysis for shale gas identification.The analysis of the high-level features extracted from the model shows that the proposed model can accurately characterize the shale gas distribution.The research contents and results of this paper are as follows:1.Based on common deep learning networks,this paper research including Convolutional Neural Networks,Recurrent Neural Networks,Autoencoder,Deep Belief Networks,and more.Research deep learning parameter optimization algorithm and feature extraction method.The key researches are two unsupervised algorithm networks of Sparse Autoencoder Network and Convolution Autoencoder Network.The framework of deep learning is built by TensorFlow,and the relevant network model is designed and established.2.According to the theory of time-frequency analysis methods for seismic signals,the research focuses on Short-time Fourier Transform,Continuous Wavelet Transform,S-transformation,Matching pursuit and Hilbert Transform.According to the theory,compare the effects of different time-frequency analysis methods on simulated seismic signals,and select the time-frequency analysis method for sparse autoencoder attribute fusion.3.Build a sparse autoencoder(SAE)network with 7 layers,and set sparse coefficient is 0.1.Using Adam optimization methods to optimize the network layer by layer.In order to achieve the purpose of extracting the fusion features for the preferred seismic attributes during the encoding process.Construct 17 layers convolutional autoencoder(CAE)network.The ReLu activation function and the Adam algorithm are used to optimize the network parameters for the gas feature extraction from the fusion features.4.Seismic data of the Wufeng-Longmaxi reservoir in the DS area of the southeastern Sichuan Basin are extracted,and use target of attribute of the short-time fourier transform,the continuous wavelet transform,the s-transform,the morlet wavelet,matching pursuit and hilbert transform for sparse autoencoder network training.The results show that the model can eliminate redundant information in attributes and enhance the gas-sensitive characteristics in the fusion data volume.5.Using the convolutional autoencoder extract the gas features from the fused feature.Extract 32 convolution features at the highest level of each fused data sample.By analyzing the correlation coefficient between each convolution feature and the target feature of the gas region,and choose better gas feature.Using the weighted average algorithm for feature,and use K-means cluster analysis to obtain the gas distribution.By comparing with the well position,the predicted gas reservoir distribution is clear and corresponds the well logging position.The research results show that the detection results based on this method are reliable.This method provides a new idea for reservoir identification.