一种基于预训练语言模型XLNet的测井曲线重构方法  

A Well Logging Curve Reconstruction Method Based on Pre-trained Language Model XLNet

作　　者：曹茂俊[1] 赵宇杰 CAO Mao-jun;ZHAO Yu-jie(School of Computer and Information Technology,Northeast Petroleum University,Daqing 163318,China)

机构地区：[1]东北石油大学计算机与信息技术学院,黑龙江大庆163318

出　　处：《计算机技术与发展》2025年第2期183-190,共8页Computer Technology and Development

基　　金：东北石油大学特色领域团队专项项目(2022TSTD-03);黑龙江省建设项目(YJSKCSZ_202309,SJGY20220253);东北石油大学建设项目(YJSXQGJKC_202204,YJSJPKC_202304)。

摘　　要：在油田勘探开发过程中,测井曲线作为地球物理测井的第一手资料,能够真实反映地下空间的分布与特性。然而,在实际工作中,由于井壁垮塌和仪器故障等原因,部分测井数据常常出现失真或缺失。为解决这一问题,该文提出了一种基于预训练语言模型XLNet的测井曲线重构方法。该方法通过筛选地层地质岩性特征指数,获取高质量的训练样本,并将其作为预训练模型重构测井曲线的依据。构建并训练带有预训练权重信息的XLNet模型,使模型具备对复杂地层特性的理解和数据重构能力。在模型的构建与训练过程中,引入了预训练权重,并进一步结合了LoRA(Low-Rank Adaptation)模块,以充分利用测井曲线之间的高度依赖关系,进而辅助XLNet生成和补全失真或缺失的测井数据。与已知曲线重构模型:基于注意力表征的长短期记忆神经网络(LSTM-Attent)、双向门控神经网络(BiGRU)、TimesNet及XLNet相比,基于预训练语言模型XLNet-LoRA的测井曲线重构模型具有更高的预测准确性。In the process of oilfield exploration and development,the logging curve,as the first-hand data of geophysical logging,can truly reflect the distribution and characteristics of underground space.However,in practice,due to borehole collapse and instrument failure,some logging data are often distorted or missing.In order to solve this problem,we propose a logging curve reconstruction method based on the pre-trained language model XLNet.In this method,high-quality training samples are obtained by screening the formation geological lithology characteristic index,which is used as the basis for the pre-training model to reconstruct the logging curve.An XLNet model with pre-trained weight information is constructed and trained,so that the model has the ability to understand the characteristics of complex strata and reconstruct data.In the process of model construction and training,we introduce pre-training weights and further combine LoRA(Low-Rank Adaptation)module to make full use of the high dependence between logging curves,and then assist XLNet to generate and complete distorted or missing logging data.The curve reconstruction model based on the pre-trained language model XLNet-LoRA has higher prediction accuracy than the known logging curve reconstruction models:long short-term memory neural network based on attention representation(LSTM-Attent),bidirectional gating neural network(BiGRU),TimesNet and XLNet.

关 键 词：测井曲线重构 深度学习 预训练语言模型 XLNet网络 LoRA机制 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]