| 融合物理约束与对抗网络的核磁测井生成方法 |
投稿时间：2026-01-17 修订日期：2026-01-28  点此下载全文 |
| 引用本文： |
| 摘要点击次数: 4 |
| 全文下载次数: 0 |
|
| 基金项目:黑龙江省科技创新基地项目“数智化油田信息感知与智能分析处理关键技术研究”（JD24A009）；东北石油大学人才引进科研启动经费资助项目（项目编号：13051202402） |
|
| 中文摘要:针对核磁共振测井在油气储层评价及勘探开发决策中存在数量稀缺及获取成本高的主要困境，本文提出一种融合物理约束与对抗网络的生成方法，在生成器中嵌入曲线-特征双端输入、地质物理约束及稳态残差网络，并在判别器的损失函数中融合物理损失。首先通过稳态网络预训练以捕获图谱宏观特征，首先通过稳态网络训练让模型预先学习图谱初始特征，再经过通过真实物理标签的重构损失，强制模型输出符合地质规律的图谱隐表示，最后经由残差网络对图谱进行精修。再引入物理约束与残差网络机制实现精细化重构。实验结果表明，在引入稳态残差网络与地质物理约束模块后，图像多样性指标为47.661，结构相似性指标为0.882，相较于其他模型的平均精度均值达到52.29%。 |
| 中文关键词:物理信息约束 生成对抗网络 核磁共振测井 残差网络 |
| |
| A method for generating nuclear magnetic logging based on integrating physical information and adversarial networks |
|
|
| Abstract:In response to the main challenges of scarce quantity and high acquisition cost in nuclear magnetic resonance logging for evaluating oil and gas reservoirs and making exploration and development decisions, this paper proposes a generation method that integrates physical constraints and adversarial networks. The generator is embedded with dual-end input of curves and features, geological-physical constraints, and steady-state residual networks, and the loss function of the discriminator is combined with physical losses. Firstly, the steady-state network is pre-trained to capture the macroscopic features of the graph, and the model is first trained through the steady-state network to learn the initial features of the graph. Then, through the reconstruction loss based on the real physical labels, the model is forced to output the graph latent representation that conforms to geological laws. Finally, the graph is refined through the residual network. The introduction of physical constraints and the residual network mechanism realizes refined reconstruction. Experimental results show that after introducing the steady-state residual network and the geological-physical constraint module, the image diversity index is 47.661, the structural similarity index is 0.882, and the average accuracy mean compared to other models reaches 52.29%. |
| keywords:physical information constraints Generative adversarial network Nuclear magnetic resonance logging Residual network |
| 查看全文 查看/发表评论 下载pdf阅读器 |
位访问者,今日访问
