Research On Fracture-vug Identification And Joint Multiwave Inversion Based On Machine Learning

With the rapid development of artificial intelligence and machine learning,machine learning method has penetrated into every link of petroleum exploration and development.Machine learning methods have penetrated into all aspects of petroleum exploration and development,and have had an important impact on petroleum geophysical exploration.They have also brought new opportunities and breakthroughs for petroleum geophysical exploration.Petroleum geophysical exploration,especially well logging and seismic exploration,often encounters a series of classification and regression problems during the research process.Based on the investigation and summary of the classification and regression problems in petroleum geophysical exploration,this paper studies the classification problem of fracture-vug identification and the regression problem of joint multiwave inversion,and presents the improved methods.For the classification problem of fracture-vug identification,a new identification method based on least squares support vector machine(LSSVM)is proposed in this paper.With FMI logging image and core observation data,the fracture and fracture-vug fillings are classified.By analyzing the log response characteristics to fracture and fracture-vug fillings,sensitive logging curves are picked out from numerous conventional logging curves.The effect of identification by using single log curve is often not good,so the Composite parameters which are more sensitive to fracture and fracture-vug fillings are extracted.LSSVM is used to establish the identification model for fracture and fracture-vug fillings.In order to improve the recognition accuracy,Particle swarm optimization(PSO)is used to optimize the parameters of LSSVM.In actual data testing,the recognition accuracy of proposed method is higher than that of BP neural network method.The recognition results are in good agreement with logging image and core data,which proves the feasibility and applicability of proposed identification method.For joint multi-wave inversion,a non-linear inversion method based on improved Bayesian inference and LSSVM is proposed in this paper.The exact Zoeppritz equation is employed in multiwave forward modeling to avoid the error of approximations due to far offset or violent variation of elastic parameters.In order to solve the non-linear problem of multi-wave joint AVO inversion,a LSSVM nonlinear model is established to define the relationship between the reflection amplitude of PP-wave,PS-wave and the elastic parameters.The improved Bayesian inference is used to determine the optimal LSSVM hyperparameters by maximizing the posterior probability of the hyperparameters.The accuracy of multi-wave joint inversion is improved by applying optimal hyperparameters.The superiority and anti-noise of the proposed method are proved by testing on synthetic data.Compared with the conventional method,the inversion results of the proposed method is more consistent with the actual logging curve and the inversion error is smaller in actual data testing.It shows that the proposed method has strong feasibility and applicability to perform the nonlinear joint inversion of the actual multi-wave seismic data in the study area.