Research On Logging Method Based On Generative Model And Attention Mechanism

In order to take the nation’s energy lifeline into its own hands,the country is vigorously developing logging technologies related to the exploration and development of fossil energy.How to accurately and efficiently grasp the underground reservoir,lithology,logging curves and other logging techniques has become the key to geological exploration.Geological exploration is very time-consuming and laborious and existing logging tools cannot meet the measurement needs under complex geological and physical conditions.Therefore,mining existing logging data by machine learning to understand the subsurface profile has a crucial role in the field of oil and gas exploration.There are still three problems that need to be solved by the current machine learning approach.Firstly,due to the complex natural environment,the logging curve data collected by sensors can have certain outliers or even missing;secondly,there is the problem that the machine learning methods currently used for lithology identification tasks do not have high generalization ability on neighboring blind wells;finally,the logging staff does not have sufficient theoretical knowledge of machine learning and still has some difficulty in using machine learning models,thus how to lower the threshold of using machine learning and make the latest technology of logging more popular is also an urgent problem to be solved.To address these problems,the following aspects are explored in this paper.Firstly,for the task of missing logging curve data completion,this thesis proposes an LSTM-ATTN model based on the long short-term memory model(LSTM)and the attention mechanism,which obtains the influence weight of each depth point on the final prediction result through the attention mechanism to better obtain the correlation of the logging curve data in the depth dimension,and then achieves the efficient completion of missing logging curve data.Secondly,for the task of reservoir lithology identification and classification,the CVAE-CGAN architecture based on the idea of generative model is proposed,which learns to fit the probability distribution of logging data,samples the characteristic relationship of logging data from logging data,and efficiently learns the correlation of nonlinear relationship between different logging curve data to achieve accurate identification and classification of reservoir lithology.Finally,a desktop client software was designed and developed based on Python language using Pyqt5 and scikit-learn libraries in order to make the logging automation methods easily available to loggers.This is an attempt to develop a domestic industrial software for logging and to popularize logging technology.