摘要
测井曲线记录着地层物理性质随深度变化的幅值范围,是测井与地震资料之间的纽带,对储层岩性分析与识别和后续的油气勘探工程十分重要.然而,在实际的测井过程中仪器故障等原因会造成测井曲线缺失的问题,重新测井不仅价格昂贵而且难以实现.针对地质勘探时测井数据时常缺失的问题,本文提出了一种LSTM(Long Short-Term Memory)注意力表征的测井曲线重构方法.同时,对原始测井信号进行两种模态分解,计算分解后得到模态分量与原始信号之间的相关性,去除冗余分量,实现对缺失的测井曲线高效、高精度的人工补全.将该方法用于声波(ACoustic,AC)与密度(DENsity,DEN)曲线重构实验,并将实验结果与LSTM网络和BP(Back Propagation)神经网络预测的结果进行对比分析.结果表明,LSTM-Attention模型有着更为优异的预测效果,重构后的AC和DEN与原始曲线之间的相关性分别达到了86.8%和74.8%,高于传统LSTM和BP神经网络预测方法.在去除冗余的信号分量后,相关系数分别提高了1.4%和4.0%.同时,本文所提方法预测出的测井曲线具有最低的预测误差.因此,基于LSTM注意表征的网络结构对测井曲线重构具有较好的预测精度.
The well logging curve records the amplitude range of geophysical properties changing with depth and is the bond between well log and seismic data.It is also significant for reservoir lithology analysis and subsequent oil and gas exploration projects.However,instrument failure and other reasons will cause well-logging curves to be missing in the actu⁃al logging process.Re-logging is not only expensive but also difficult to achieve.Aiming at the problem that logging data is often missing during geological exploration,this paper proposes a logging curve reconstruction method based on the LSTM(Long Short-Term Memory)-attention model.At the same time,EMD-VMD(Empirical Mode Decomposition-Variational Mode Decomposition)decomposition is performed on the original logging signal and then the correlation between the com⁃ponents and the original curve is calculated.Some excrescent components are deletedto promote efficient and high-preci⁃sion manual completion.This proposed method is applied to reconstruct missing logs acoustic(AC)and density(DEN),and the prediction results are compared with those predicted by LSTM and BP(Back Propagation)neural network.The results show that the LSTM-attention model has a better prediction performance,and the correlations between predictive and the original curves can reach 86.8%(AC)and 74.8%(DEN),higher than the traditional LSTM and BP neural network.After removing redundant signal components,the correlation coefficients increased by 1.4%(AC)and 4.0%(DEN).At the same time,the logging curve predicted by the proposed method has the lowest prediction error.Therefore,the representation learning based on LSTM with an attention mechanism has better prediction accuracy for well-logging curve reconstruction.
作者
刘梦
韩建
曹志民
刘兴斌
LIU Meng;HAN Jian;CAO Zhi-min;LIU Xing-bin(School of Physics and Electronic Engineering,Northeast Petroleum University,Daqing,Heilongjiang 163319,China;Research and Development Center for Testing and Measurement Technology and Instrumentation,Heilongjiang Province University,Northeast Petroleum University,Daqing,Heilongjiang 163319,China)
出处
《电子学报》
EI
CAS
CSCD
北大核心
2024年第4期1399-1410,共12页
Acta Electronica Sinica
基金
国家自然科学基金(No.52174021)。
