Scholarlink

<p>This paper proposes a machine learning technique based on “training while drilling” to improve the integrity and accuracy of resistivity logging data in the oil and gas industry. Resistivity logging is a critical tool for assessing hydrocarbon content in geological formations; however, it often suffers from data loss due to hardware failures caused by the harsh high-temperature conditions encountered downhole. To address these challenges, this study introduces Long Short-Term Memory (LSTM) neural networks, a deep learning technique well-suited for handling time-series data and predicting missing values. Unlike traditional surface-based data processing, this method integrates the trained LSTM model into embedded devices, enabling real-time downhole data completion and significantly enhancing the efficiency and immediacy of data processing. “training while drilling” refers to the continuous learning and model updating based on newly acquired data during drilling operations. This process utilizes downtime, such as during drill pipe or bit changes, to input fresh data as a training set, allowing for real-time model optimization. Experimental results demonstrate that this method notably improves the predictive accuracy of the LSTM model, with enhanced capability to capture detailed information and adapt to changes in formation characteristics when compared to the initially trained model. Moreover, the LSTM model outperforms other deep learning algorithms, including Fully Connected Neural Networks (FCNN), in terms of prediction accuracy and stability. The dataset used in this study is based on field data from the Shengli Oilfield in Dongying, with measurements spanning depths of 1140 to 1690&#xa0;m. Through comparative experiments, the LSTM model achieved a mean squared error (MSE) of 0.0610, significantly lower than the MSE of 0.0961 observed for the FCNN and other traditional methods. These findings highlight the superior performance of the “training while drilling” approach in improving prediction accuracy of electromagnetic wave logging curves, providing a novel solution with considerable practical value and broad potential for future development in the oil and gas exploration sector.</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 method for training while drilling to predict electromagnetic wave logging curves based on long short-term memory neural networks

  • 期刊论文

  • 发表日期：

    2025年6月4日

查看全文
  • Qiwei Liu,
  • Fanmin Kong,
  • Kang Li

摘要

This paper proposes a machine learning technique based on “training while drilling” to improve the integrity and accuracy of resistivity logging data in the oil and gas industry. Resistivity logging is a critical tool for assessing hydrocarbon content in geological formations; however, it often suffers from data loss due to hardware failures caused by the harsh high-temperature conditions encountered downhole. To address these challenges, this study introduces Long Short-Term Memory (LSTM) neural networks, a deep learning technique well-suited for handling time-series data and predicting missing values. Unlike traditional surface-based data processing, this method integrates the trained LSTM model into embedded devices, enabling real-time downhole data completion and significantly enhancing the efficiency and immediacy of data processing. “training while drilling” refers to the continuous learning and model updating based on newly acquired data during drilling operations. This process utilizes downtime, such as during drill pipe or bit changes, to input fresh data as a training set, allowing for real-time model optimization. Experimental results demonstrate that this method notably improves the predictive accuracy of the LSTM model, with enhanced capability to capture detailed information and adapt to changes in formation characteristics when compared to the initially trained model. Moreover, the LSTM model outperforms other deep learning algorithms, including Fully Connected Neural Networks (FCNN), in terms of prediction accuracy and stability. The dataset used in this study is based on field data from the Shengli Oilfield in Dongying, with measurements spanning depths of 1140 to 1690 m. Through comparative experiments, the LSTM model achieved a mean squared error (MSE) of 0.0610, significantly lower than the MSE of 0.0961 observed for the FCNN and other traditional methods. These findings highlight the superior performance of the “training while drilling” approach in improving prediction accuracy of electromagnetic wave logging curves, providing a novel solution with considerable practical value and broad potential for future development in the oil and gas exploration sector.