测井技术2025，Vol.49Issue(6)：857-867,11.DOI:10.16489/j.issn.1004-1338.2025.06.004
基于粒子群优化Transformer模型的声波测井曲线重构方法
Method for Sonic Log Reconstruction Based on a Particle Swarm-Optimized Transformer Model
摘要
Abstract
The sonic log(AC)plays a critical role in evaluating formation porosity and rock mechanical properties.However,during actual logging operations,issues such as equipment failure or complex downhole conditions often lead to missing or severely distorted log segments.Therefore,the effective reconstruction of sonic logs is necessary.This study employs an integrated approach based on petrophysical models and correlation analysis(using Pearson and Spearman coefficients)to identify the optimal sensitive variables for AC reconstruction:Gamma ray(GR),deep laterolog resistivity(RLLD),and compensated neutron log(CNL).Subsequently,a deep learning Transformer model is utilized to capture the spatio-temporal correlations among these multiple log variables through its multi-head attention mechanism and autoregressive structure.The particle swarm optimization(PSO)algorithm is introduced to perform a global search and fine-tune the model's hyperparameters(e.g.,learning rate,number of attention heads,hidden layer dimensions),thereby enhancing the model's prediction accuracy and stability for precisely reconstructing missing or distorted AC curves.Comparative experiments are conducted against multiple regression and fully connected neural network(FNN)models.The results demonstrate that:①Based on correlation analysis and petrophysical model evaluation,GR,RLLD,and CNL are identified as the optimal sensitive variables for AC reconstruction.The Pearson correlation coefficient between AC and CNL reaches 0.88,and the absolute values of the Spearman correlation coefficients between AC and GR/RLLD both exceed 0.60,effectively ensuring input data validity.②The Transformer model accurately captures the complex nonlinear relationships among the logging variables,and the application of the PSO algorithm further enhances its predictive capability,culminating in the development of the PSO-Transformer based AC reconstruction method.③Comparative experiments show that the PSO-Transformer model achieves the best reconstruction performance,with a relative error controlled within 5%.On the test set from well A1,it achieved a mean square error(IMSE)of 10.914 0,a mean relative error(IMRE)of 0.028 8,and a coefficient of determination(R2)of 0.901 6,surpassing the multiple regression model(IMSE:19.805 7,IMRE:0.064 3,R2:0.823 7)and the FNN model(IMSE:16.403 6,IMRE:0.038 5,R2:0.857 8).Furthermore,the reconstructed logs maintain strong continuity with the original AC curve.④The model is suitable for sonic log reconstruction under complex geological conditions,although corrections based on regional geology are recommended for severely enlarged borehole sections.In conclusion,this study confirms that the PSO-Transformer model can efficiently and accurately reconstruct missing or distorted AC logs,outperforming traditional multiple regression and FNN models.It provides reliable log data support for formation evaluation and integrated geological engineering applications in complex geological settings.关键词
曲线重构/声波测井/Transformer模型/粒子群优化算法/地质工程一体化/深度学习Key words
log reconstruction/acoustic logging/Transformer model/particle swarm optimization algorithm/integrated geological and engineering evaluation/deep learning分类
天文与地球科学引用本文复制引用
WANG Guohao,YAN Jianping,QIU Xiaoxue,LIAO Maojie,YANG Yang,YAN Hua..基于粒子群优化Transformer模型的声波测井曲线重构方法[J].测井技术,2025,49(6):857-867,11.基金项目
国家自然科学基金项目"低电阻率页岩气储层:成因机制差异及含气饱和度模型研究"(42372177) （42372177）
中国石油-西南石油大学创新联合体科技合作项目"川南深层与昭通中浅层海相页岩气规模效益开发关键技术研究"(2020CX020000) （2020CX020000）
四川省自然科学基金项目"页岩气储层低电阻率成因机制及对含气性的影响研究"(2022NSFSC0287) （2022NSFSC0287）
