Improved Convolution Neural Networks For Stratigraphic Identification

Stratigraphic identification refers to the identification of the stratigraphic sequence within an oil zone.This is done by comprehensively processing the formation information of the entire well section.The recognition result directly affects the subsequent evaluation results of the reservoir.It has an important guiding role for oil and gas field exploration.Logging curves contain rich geological information,which is of great significance for stratigraphic identification,and is usually used as the basis for stratigraphic identification.Artificial stratigraphic identification methods are inefficient and subjective,and cannot meet the objective needs of reservoir evaluation in complex geological environments.The feature extraction method of the convolutional neural network is similar to the method of manually analyzing the logging curves,and the convolutional neural network can be used to realize the automatic stratigraphic identification.However,the calculation efficiency of traditional convolutional neural network is not ideal for complex logging curves images.This article has conducted a more systematic and in-depth study on the improvement of convolutional neural network and its application in stratigraphic identification.The main content includes the following two aspects.First,to solve the problem that the identification result declines when the resolution of the logging curve characteristic map decreases,an improved high-resolution network model is proposed.Combining the attention mechanism module and the deep separable convolution structure in the improved high-resolution network,combined with the multi-scale feature extraction advantages of the high-resolution network itself,it effectively improves the accuracy of stratum feature recognition,and the optimal nonlinear activation function makes network calculation faster and quantification is more friendly.Secondly,in view of the similarity of logging curve images and the problem that redundant feature maps are easily generated during the convolution process,an improved Ghost network model is proposed.First,the Ghost module structure is used as the basic module to solve the problem of feature image redundancy through cheap operations,and the ordinary convolution in the Ghost module is changed to expanded convolution to obtain a multi-scale representation;second,a bi-directional cascaded network is constructed and a bi-directional loss function is used for supervising the learning process.Each intermediate layer can learn its own appropriate scale to enhance the recognition accuracy of the target level;finally,build a bi-directional cascaded Ghost network model.The bi-directional supervision mechanism of each layer of the network solves the problem of training strategy.Network performance has been steadily improved.This thesis selects the logging data of an oilfield in the southwest,first conduct ablation experiments on the improved network model to obtain the best network structure,and then conduct a comprehensive comparison experiment with similar convolutional neural network models.The identification results under different training/testing dataset sample ratios and the rate of recall and precision are higher than the comparison method.The experimental results reveal that the two improved network models proposed in this thesis can better identify the stratigraphic sequence,and have certain theoretical and practical significance for the study of stratigraphic identification.