Research And Application Of Seismic Fracture-cavity Detection Method Based On Deep Learning

Carbonate reservoir is one of the main sources of petroleum reserves in China,and many carbonate oil and gas fields with a production of up to one million tons have been found in northwest China.Fracture-cavity reservoir is a typical carbonate reservoir,and its superior connectivity makes it a natural reservoir space.Therefore,the identification of fracture-cavity reservoir is of great significance to the development of carbonate reservoir.Traditional fracture-cavity identification relies too much on human calibration,which is highly subjective and multi-solvable,and is also affected by the quality of seismic data.With the continuous application of deep learning in the field of seismic interpretation,this paper introduces deep learning into fracture-cavity identification,turns karst cave identification into an image segmentation problem,and summarizes a set of fracture-cavity identification process with advanced fracture detection algorithm.To achieve intelligent,high precision can simultaneously characterize the cave and the cave target.In this paper,we propose a method for karst cave recognition based on U-net deep convolutional network.On the basis of the basic U-net convolutional neural network model,we combine the residual network module and U-net network.In order to reduce the possibility of overfitting,we add the Dropout layer for regularization.Only one 3×3×3 convolutional kernel is used for each layer of the network,so as to improve the computing power and efficiency of the network and avoid the problem of overfitting.To solve the problem of insufficient data set,the information contained in the data set is expanded by adding actual data and the data marked by geological experts,so as to reduce the rate of missing cave detection and increase the accuracy of karst cave identification.In this paper,a high-precision crack identification method is also proposed.Based on the coherence volume,through the innovation of heuristic factors,the waveform is decomposed into crack scale,strike and geometric shape,and edge detection is added.The heuristic function is established by multi-factor convolution for ant tracking,and finally the crack distribution data including crack strike and dip Angle information is obtained.After obtaining the data bodies representing the karst cave and crack respectively,the two attributes are fused through the depth feedforward neural network,and the data which can be used for the identification of the cracks is output.After testing the method,the actual data of an oilfield in Tarim Basin are applied.The results show that the proposed method has high accuracy and has certain guiding significance for actual production.