Scholarlink

<p>Lithology identification using well log data is a crucial part of geophysical reservoir characterization. However, lithologies have different proportions in geological formations, which means that an imbalance exists in the well log datasets. When confronted with this issue, the performance of intelligent lithology identification models will be constrained. To address the issue, this paper proposes a lithology identification framework based on a denoising diffusion probabilistic model (DDPM) and a multiscale convolutional neural network (MCNN) for imbalanced well log data. The DDPM is designed to generate minority category well log data to mitigate the impact of dataset imbalance. These generated samples are then evaluated by a discriminator and filter module to select high-quality samples, which are subsequently integrated into the original dataset for balancing. The MCNN then identifies these balanced data using multiscale convolution and attention mechanisms. Experiments are conducted on two datasets from the Hugoton and Panoma fields in Kansas, USA, and the Daqing fields in China. The DDPM-MCNN achieves recognition accuracy of 84.99% and 85.58%, respectively, offering a novel approach for lithology identification on imbalanced well log datasets.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Lithology Identification of Imbalanced Well Log Data Based on Diffusion Model and Multiscale CNN

  • 期刊论文

  • 发表日期：

    2025年3月22日

查看全文
  • Fengda Zhao,
  • Zhuoyi Zhao,
  • Hongjin Lv,
  • Pengwei Zhang,
  • Xianshan Li

摘要

Lithology identification using well log data is a crucial part of geophysical reservoir characterization. However, lithologies have different proportions in geological formations, which means that an imbalance exists in the well log datasets. When confronted with this issue, the performance of intelligent lithology identification models will be constrained. To address the issue, this paper proposes a lithology identification framework based on a denoising diffusion probabilistic model (DDPM) and a multiscale convolutional neural network (MCNN) for imbalanced well log data. The DDPM is designed to generate minority category well log data to mitigate the impact of dataset imbalance. These generated samples are then evaluated by a discriminator and filter module to select high-quality samples, which are subsequently integrated into the original dataset for balancing. The MCNN then identifies these balanced data using multiscale convolution and attention mechanisms. Experiments are conducted on two datasets from the Hugoton and Panoma fields in Kansas, USA, and the Daqing fields in China. The DDPM-MCNN achieves recognition accuracy of 84.99% and 85.58%, respectively, offering a novel approach for lithology identification on imbalanced well log datasets.