Lithology Identification Of Volcanic Rock Reservoirs Based On Deep Learning

Due to the complexity of volcanic rock lithology,the conventional identification methods for lithology are limited by the feature extraction capability,resulting in poor application effectiveness.Deep learning algorithms have excellent feature extraction capability and can extract information that low-dimensional analysis cannot extract from input data.Their outstanding performance has played a powerful role in multiple fields.In order to improve the effectiveness of lithology identification in the study area and achieve the accuracy required for practical production,this paper conducted in-depth research on several deep learning algorithms applied to the identification of lithology in logging volcanic rock formations.The volcanic rock lithology identification research in this paper is based on conventional logging data.In order to ensure the accuracy of sample data,a sample data set was selected from well segments with core debris in block 31,and the core conclusion was used as the sample label.The logging data and core debris data in the study area were used for stratigraphic feature research and typical logging response feature analysis to fully understand the geological overview of the study area.Based on the analysis of typical logging response features,the conventional lithology identification methods based on the box-and-whisker plot method and crossplot method were conducted.However,due to the fact that conventional identification methods can only classify some lithology categories,the identification effect is poor,and there are still many different categories of lithology interwoven together,which makes the accuracy of lithology identification lower than the requirements for practical production.In order to improve the accuracy of lithology identification in the study area,a bidirectional long short-term memory neural network based on a recurrent neural network was proposed to take the correlation of logging curves into account as the current stratigraphic lithology identification method.Compared with machine learning algorithms such as random forest and multilayer perceptron,this method can input logging data as a sequence into the model,extract the correlation information of logging data in depth,and identify stratigraphic lithology by synthesizing the feature information before and after that depth.Transformer is a neural network that uses attention mechanisms,which can focus more on key features.Therefore,its performance in many fields is better than that of recurrent neural networks and convolutional neural networks.Based on the characteristics of Transformer,an Informer lithology identification method is proposed.Before conducting the deep learning lithology identification study,the logging data was normalized and the curves were optimized.This paper selected the optimal logging curves through correlation analysis and factor analysis.The selected logging curves were then sampled at a 0.125-meter interval and grouped into eight intervals through interpolation.To extract the deep correlation features of the logging curves,a study on lithology identification based on bidirectional long short-term memory neural networks was conducted.The results showed that the bidirectional long short-term memory neural network had good performance in lithology identification,with an accuracy of 81.2%when the model batch size was set to 64,the iteration number was set to 2000,and the dropout rate was set to 0.4.To further extract the deep correlation features of the logging data,attention mechanisms were introduced,and a lithology identification study based on Informer was conducted,which preserved the correlation of the logging data in depth using position encoding techniques.The study showed that Informer had a higher recognition accuracy than the bidirectional long short-term memory network,with an accuracy of 86.2% when the number of multi-head attention was set to 16 and the convolutional layer size was set to 64.Compared with the bidirectional long short-term memory neural network,the accuracy was improved by nearly 5%.