测井技术2025，Vol.49Issue(6)：879-889,11.DOI:10.16489/j.issn.1004-1338.2025.06.006
基于小波变换与CMT架构融合的地层智能划分与对比研究
Research on Intelligent Stratigraphic Division and Correlation with Integrated Wavelet Transform and CMT Architecture
摘要
Abstract
To address the issues of traditional stratigraphic correlation methods,which rely on manual identification of marker beds and sedimentary cycles,and suffer from heavy workload,low correlation accuracy,and high result uncertainty,as well as to improve the accuracy and efficiency of automatic well log stratification,an automatic stratigraphic division and correlation method integrating wavelet transform and the CNN(convolutional neural network)meet Transformer(CMT)architecture is adopted.Taking the Chang7 and Chang8 intervals in the Dadonggou block of Fuxian county as the research object,this method performed five-level wavelet decomposition on gamma-ray well logs to extract approximation components and detail components,combined GR and acoustic travel time curves as multi-channel inputs,extracted local stratigraphic features via the CNN module,modeled global sequence dependencies through the Transformer module,and realized automatic stratification through three stages:feature extraction,training,and prediction.Experiments with different training data scales(60%,40%,20%)and comparisons with multiple models(SegNet,CNN,Transformer)are also designed.The research results show that:① With 60%training data,the model's average accuracy reaches 0.882 5,and the mean absolute error between the predicted and actual small-layer boundary depths is 2.432 4 m;with 40%training data,the average accuracy remains 0.855 0,and the mean absolute error is 3.124 0 m;while with 20%training data,the accuracy decreases significantly(accuracy 0.821 2,mean absolute error is 3.886 1 m).② After introducing wavelet transform features,the accuracy of the CMT model increases from 0.816 9 to 0.850 8,and the mean absolute error decreases from 3.965 2 m to 3.012 7 m.③ Under the condition of 40%training data,the average accuracy of the CMT model is significantly better than that of SegNet,CNN and Transformer,especially performing better in thin-layer identification and unbalanced sample scenarios.④ A training data volume of 40%can meet the feasibility of engineering applications and reduce human and material resource consumption.It is concluded that the CMT method integrated with wavelet transform effectively overcomes the limitations of traditional stratification methods,such as strong subjectivity and low efficiency.Wavelet transform enhances the model's ability to identify geological cycles and sequence boundaries,and the CMT architecture fully combines the advantages of CNN in local feature extraction and Transformer in global modeling,providing an intelligent and refined technical approach for stratigraphic correlation and reservoir modeling.关键词
CMT(CNN Meet Transformer)/地层智能划分/地层对比/测井曲线/小波变换/自动分层/自然伽马/沉积旋回Key words
CNN meet Transformer(CMT)/intelligent stratigraphic division/stratigraphic correlation/well logging curve/wavelet transform/automatic stratigraphic division/gamma-ray/sedimentary cycle分类
天文与地球科学引用本文复制引用
ZHANG Gang,WANG Junhui,HUI Xin,AN Shan,PENG Qiangqiang,WAN Lei..基于小波变换与CMT架构融合的地层智能划分与对比研究[J].测井技术,2025,49(6):879-889,11.基金项目
国家自然科学基金项目"基于扩散模型的迷宫状油气储层建模方法研究"(42472209) （42472209）
