Scholarlink

<p>Accurate lithology detection is pivotal in the oil and gas industry, offering substantial cost-saving potential and the elimination of the need for expert field knowledge. This study explores automated lithology detection using well logs, focusing on a dataset of wire-line log measurements from the Hugoton field in Kansas, encompassing nine facies classes: Nonmarine sandstone (SS), Nonmarine coarse siltstone (CSiS), Nonmarine fine siltstone (FSiS), Marine siltstone and shale (SiSh), Mudstone (MS), Wackestone (WS), Dolomite (D), Packstone-grainstone (PS), and Bafflestone (BS). We employ a sophisticated preprocessing technique, inspired by human activity recognition literature, to transform well log data into signal images (SIs). These images are classified using deep learning (DL) models of varying capacity, including single-layer, double-layer, and VGG-16 networks, to recognize facies. Rigorous evaluations assess model performance and robustness under common variations and limitations. The integration of advanced data processing and DL feature extraction holds the potential for achieving 100% accuracy, significantly surpassing the performance of previously reported alternatives for the same dataset in the literature. We employ additional metrics such as F1-score, recall, and precision to comprehensively analyze model performance. This achievement enables cost-effective, data-driven lithology recognition strategies. The best performance in our approach is achieved when computed features from the raw data are augmented in model training. Additionally, we perturb the data with Gaussian noise of different levels to evaluate model’s performance under less-than-ideal conditions. Understanding how well the deep models perform in the presence of noise enhances the study’s practicality, ensuring the proposed lithology detection system remains accurate and reliable, even with minor data imperfections.</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

Transforming petrophysical well-logs to images: leveraging deep learning for lithology recognition

  • 期刊论文

  • 发表日期：

    2025年7月1日

查看全文
  • Kamyar Mohammadi,
  • Seyed Hamid Reza Mousavi,
  • Seyed Mojtaba Hosseini-Nasab,
  • Mohammad Shahbazi

摘要

Accurate lithology detection is pivotal in the oil and gas industry, offering substantial cost-saving potential and the elimination of the need for expert field knowledge. This study explores automated lithology detection using well logs, focusing on a dataset of wire-line log measurements from the Hugoton field in Kansas, encompassing nine facies classes: Nonmarine sandstone (SS), Nonmarine coarse siltstone (CSiS), Nonmarine fine siltstone (FSiS), Marine siltstone and shale (SiSh), Mudstone (MS), Wackestone (WS), Dolomite (D), Packstone-grainstone (PS), and Bafflestone (BS). We employ a sophisticated preprocessing technique, inspired by human activity recognition literature, to transform well log data into signal images (SIs). These images are classified using deep learning (DL) models of varying capacity, including single-layer, double-layer, and VGG-16 networks, to recognize facies. Rigorous evaluations assess model performance and robustness under common variations and limitations. The integration of advanced data processing and DL feature extraction holds the potential for achieving 100% accuracy, significantly surpassing the performance of previously reported alternatives for the same dataset in the literature. We employ additional metrics such as F1-score, recall, and precision to comprehensively analyze model performance. This achievement enables cost-effective, data-driven lithology recognition strategies. The best performance in our approach is achieved when computed features from the raw data are augmented in model training. Additionally, we perturb the data with Gaussian noise of different levels to evaluate model’s performance under less-than-ideal conditions. Understanding how well the deep models perform in the presence of noise enhances the study’s practicality, ensuring the proposed lithology detection system remains accurate and reliable, even with minor data imperfections.