Scholarlink

<p>Logging-while-drilling (LWD) azimuthal electromagnetic resistivity tools can obtain the amplitude ratios and phase differences during drilling, but these signals cannot directly quantify the formation resistivity or boundary distances which are essential for geosteering. Logging data inversion is required to extract these parameters, there are two challenges: traditional iterative methods have low efficiency, whereas deep learning models lack interpretability. Thus, this study introduces an interpretable deep learning framework to solve the complex problem of real-time LWD inversion for the first time. This framework integrates the residual next network (ResNeXt) feature extraction mechanism with the architecture of self-explaining neural network (SENN) to resolve the issue of opaque decision-making in black-box models. Using multi-task learning principles and masking mechanisms, the framework also enables a high-precision parallel inversion of the formation interface distance and resistivity parameters. Furthermore, a concept activation index (CAI) is designed to help quantify the reliability of the model predictions. The results demonstrate that the proposed explainable model achieves excellent inversion performance. Across all parameter prediction tasks (for the formation resistivity and boundary distances), the proposed explainable model has a 20.8% lower average prediction error than the optimal models among recent LWD inversion networks and a contemporary interpretable network. The single-point inference time of the network is 0.04&#xa0;s, which satisfies real-time requirements; the CAI effectively identifies regions with unreliable predictions and provides a reliable basis for manual review and post-processing.</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

Inversion of logging data in azimuthal electromagnetic wave resistivity while drilling based on self-explaining deep learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月7日

查看全文
  • Yufeng Han

摘要

Logging-while-drilling (LWD) azimuthal electromagnetic resistivity tools can obtain the amplitude ratios and phase differences during drilling, but these signals cannot directly quantify the formation resistivity or boundary distances which are essential for geosteering. Logging data inversion is required to extract these parameters, there are two challenges: traditional iterative methods have low efficiency, whereas deep learning models lack interpretability. Thus, this study introduces an interpretable deep learning framework to solve the complex problem of real-time LWD inversion for the first time. This framework integrates the residual next network (ResNeXt) feature extraction mechanism with the architecture of self-explaining neural network (SENN) to resolve the issue of opaque decision-making in black-box models. Using multi-task learning principles and masking mechanisms, the framework also enables a high-precision parallel inversion of the formation interface distance and resistivity parameters. Furthermore, a concept activation index (CAI) is designed to help quantify the reliability of the model predictions. The results demonstrate that the proposed explainable model achieves excellent inversion performance. Across all parameter prediction tasks (for the formation resistivity and boundary distances), the proposed explainable model has a 20.8% lower average prediction error than the optimal models among recent LWD inversion networks and a contemporary interpretable network. The single-point inference time of the network is 0.04 s, which satisfies real-time requirements; the CAI effectively identifies regions with unreliable predictions and provides a reliable basis for manual review and post-processing.