Scholarlink

<p>Lithology identification plays a critical role in oil and gas exploration. However, differences in sedimentary environments and instrument drift lead to significant domain shift between different wells, making it difficult for cross-well lithology identification models to maintain reliable performance on unlabeled target wells. Existing unsupervised domain adaptation methods usually only perform marginal distribution alignment, which tends to cause category mismatch in the absence of category semantic constraints. Meanwhile, a large number of low-confidence target samples are not effectively utilized, which limits the learning of target-domain structural information. To address the above problems, this study proposes a Self-training based Conditional Alignment Domain Adaptation method (STCA-DA). This method takes lithology category prototypes as conditional constraints to guide category-level alignment between the source domain and the target domain, updates the prototype structure robustly through self-training strategy and min-max entropy loss, and constructs negative samples for contrastive learning using low-confidence samples to strengthen category boundaries. Experiments on two datasets, the Hugoton and Panoma Fields Oil Fields in the USA and the Daqing Oilfield in China, show that STCA-DA significantly outperforms various baseline models, achieving maximum improvements of 5.3% and 4.3% in accuracy and <i>F</i>1-score, respectively, demonstrating superior cross-well transferability.</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

A Self-Training Conditional Alignment Framework for Cross-Well Lithology Identification

  • 期刊论文

  • 发表日期：

    2026年7月5日

查看全文
  • Xianshan Li,
  • Zhaoze Jiang,
  • Yonghang Yu,
  • Xingyun Wei,
  • Pengwei Zhang,
  • Hongjia Ren,
  • Maoyuan Sun,
  • Tianqi Li,
  • Fangfang Pan,
  • Fengda Zhao

摘要

Lithology identification plays a critical role in oil and gas exploration. However, differences in sedimentary environments and instrument drift lead to significant domain shift between different wells, making it difficult for cross-well lithology identification models to maintain reliable performance on unlabeled target wells. Existing unsupervised domain adaptation methods usually only perform marginal distribution alignment, which tends to cause category mismatch in the absence of category semantic constraints. Meanwhile, a large number of low-confidence target samples are not effectively utilized, which limits the learning of target-domain structural information. To address the above problems, this study proposes a Self-training based Conditional Alignment Domain Adaptation method (STCA-DA). This method takes lithology category prototypes as conditional constraints to guide category-level alignment between the source domain and the target domain, updates the prototype structure robustly through self-training strategy and min-max entropy loss, and constructs negative samples for contrastive learning using low-confidence samples to strengthen category boundaries. Experiments on two datasets, the Hugoton and Panoma Fields Oil Fields in the USA and the Daqing Oilfield in China, show that STCA-DA significantly outperforms various baseline models, achieving maximum improvements of 5.3% and 4.3% in accuracy and F1-score, respectively, demonstrating superior cross-well transferability.