Lithology Identification Based On Multi-Source Domain Adaptive And Semi-Supervised Generative Adversarial Network

Lithology identification is a process of recognizing and distinguishing lithology from logging curves by specific methods.It is a basic task in the fields of formation evaluation,reservoir description and geological modeling.In recent years,the rapid development of artificial intelligence technology has been widely used in the field of oil exploration,the lithology identification method based on machine learning and deep learning has gradually replaced the traditional methods relying on expert interpretation.However,there are still three main problems to be solved in the application of artificial intelligence technology in logging lithology identification.Firstly,the cross-domain problem in logging lithology identification is rarely considered in the current research,and the distribution difference of each well data is ignored,which leads to the degradation of model performance.Secondly,for the small sample logging data set composed of few logging data,the existing model cannot overcome the influence of the factors such as the small number of samples and the imbalance of distribution between classes,leading to the overfitting and identification accuracy of the model.Finally,the current artificial intelligence logging lithology identification method is still in the theoretical research stage,and it is urgent to explore the integration and engineering of artificial intelligence logging method.Therefore,how to solve the problems of poor performance and difficult application of existing models in small sample and cross-domain logging lithology identification has become the focus of research in logging field.Firstly,a multi-source domain adaptive model is proposed for the cross-domain logging lithology identification problem.The model extracts hidden information in a specific domain space by aligning multiple domains through a multi-source domain adaptive network and employs a three-stage training model to mine specific decision boundaries.The model also designs a fine-grained correction method for confusion sample correction,which employs a homogeneous maximum difference network to identify the boundaries of confusion samples to improve classification accuracy and generalization ability.Secondly,a classification-enhanced semi-supervised generative adversarial network model is proposed for the small-sample logging lithology identification problem.The model mines and extracts the nonlinear features of logging curves through a classifier network with a one-dimensional convolutional structure,and a classification task separation architecture is used to reduce the mutual interference between classification and identification tasks to ensure a balanced network convergence.Also,a pseudo-label processing mechanism is designed in the model to achieve more efficient classification enhancement by assigning pseudo-labels and supplementing supervised learning to improve recognition accuracy.Finally,an automatic logging software is designed and developed based on Python and Qt5.The software integrates a variety of machine learning and deep learning algorithms to provide users with a complete and easy to operate automated logging interpretation solution.