Research On Formation Recognition While Drilling Based On Data Augmentation And Deep Learning

Exploration and development of oil and gas resources presents a plethora of intricate formation structure types,making drilling technology highly difficult.An accurate drilling formation identification model can be employed to quickly identify the type of formation encountered,not only to pinpoint the reservoir’s position,but also to optimize the drilling process,thereby preventing downhole risks and improving drilling economics.However,the drilling formation identification is affected by many factors,including a large amount of real-time formation geological detection information,drilling process information and drilling equipment information.Therefore,thesis provides a new solution and technical means to establish a drilling formation identification method with data mining technology represented by deep learning,mainly focusing on drilling data quality,formation class imbalance and drilling formation identification accuracy and timeliness.The main work of thesis is as follows:(1)The composition and characteristics of each formation type in the Sichuan Basin are analyzed,and a large amount of drilling data is obtained through comprehensive logging techniques to determine the data basis for the analysis and study.In response to the problem of many missing values and abnormal values in drilling data,filtering and clustering methods are integrated to form a drilling data quality improvement method.On this basis,multi-source information is fully integrated,and multiple correlation algorithms are used to mine key feature parameters and time-series features related to formation response from historical sequence information,extracte the self-development pattern of each feature parameter,capture the correlation between different parameters,and eliminate irrelevant and redundant information,so as to establish a drilling formation identification index system and improve the timeliness of drilling formation identification.(2)A drilling data augmentation method with supervised learning capability is designed to address the formation class imbalance problem caused by varying formation thickness.The method is based on a conditional variational auto-encoder to learn the probability density distribution of multiple feature parameters in various stratigraphic samples,and after extracting the distribution features of the original drilling data,the real drilling data corresponding to each stratigraphic layer is reconstructed by supervised learning.The experimental results show that the data generated by this data enhancement method was highly similar to the real drilling data,which verifies the effectiveness of the method.Using this method to expand the original dataset not only solve the formation class imbalance problem and improve the generalization of the subsequent follow-on formation identification model,but also have a positive effect on the performance aspects such as model validity and accuracy.(3)A lightweight formation classification model is designed to address the problems of low accuracy and timeliness of the formation identification while drilling.The model is based on convolutional neural networks and Time Distributed(TD)framework to extract formation variation patterns in multidimensional drilling time series,and its effectiveness is verified on actual drilling data sets.Based on this,a recurrent neural network with strong temporal processing capability is selected to improve the model,replacing the upper layer of the network in the output layer of the original model.The experimental results show that the improved model is not only improved in terms of accuracy,but also more time-sensitive.