Scholarlink

<p>During complex oil and gas drilling, minor lost circulation is a common yet hard-to-monitor downhole incident, and its accurate detection facilitates the timely implementation of on-site well control measures to mitigate economic losses caused by such incidents. Two primary reasons contribute to this challenge: first, the scarcity of lost circulation data compared to normal drilling data leads to severe class imbalance; second, conventional intelligent monitoring methods fail to capture long-term data trends. To develop classification capabilities, intelligent monitoring models must learn from samples composed of data segments extracted from data sequences corresponding to normal drilling and minor lost circulation. However, the aforementioned two issues impair the models’ ability to extract and identify the feature of minor lost circulation. To address these dual challenges, this study introduces an integrated framework. First, a training sample class balancing method based on the Borderline-Synthetic Minority Oversampling Technique (Borderline-SMOTE) is proposed, which generates synthetic data specifically for samples near the decision boundary between normal drilling and minor lost circulation. Second, a novel recognition model combining Dilated Causal Convolution (DCC) and Long Short-Term Memory (LSTM) networks is developed to effectively capture long-range dependencies in sensor data. Notably, this study considers both sample class balance and model design as key factors influencing monitoring accuracy and proposes innovative solutions for each, in contrast to previous studies that focused only on one of these aspects. Experimental results on field data show that the proposed DCC-LSTM model achieves a recognition accuracy of 96.38%. When trained on the balanced dataset, six common intelligent models (k-Nearest Neighbor (kNN), decision tree, random forest, LSTM, Gated Recurrent Unit (GRU), and DCC-LSTM) exhibit significant performance improvements, with a maximum accuracy increase of 6.76%. The proposed method can be applied for minor lost circulation recognition in scenarios with limited samples, and is expected to be extended to other drilling anomaly detection tasks or complex downhole condition monitoring in the future.</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

Sample class balancing based on borderline-SMOTE for lost circulation recognition

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月23日

查看全文
  • Weifeng Sun,
  • Chentao Gong,
  • Saisai Bu,
  • Dongyue Wang,
  • Weimin Huang

摘要

During complex oil and gas drilling, minor lost circulation is a common yet hard-to-monitor downhole incident, and its accurate detection facilitates the timely implementation of on-site well control measures to mitigate economic losses caused by such incidents. Two primary reasons contribute to this challenge: first, the scarcity of lost circulation data compared to normal drilling data leads to severe class imbalance; second, conventional intelligent monitoring methods fail to capture long-term data trends. To develop classification capabilities, intelligent monitoring models must learn from samples composed of data segments extracted from data sequences corresponding to normal drilling and minor lost circulation. However, the aforementioned two issues impair the models’ ability to extract and identify the feature of minor lost circulation. To address these dual challenges, this study introduces an integrated framework. First, a training sample class balancing method based on the Borderline-Synthetic Minority Oversampling Technique (Borderline-SMOTE) is proposed, which generates synthetic data specifically for samples near the decision boundary between normal drilling and minor lost circulation. Second, a novel recognition model combining Dilated Causal Convolution (DCC) and Long Short-Term Memory (LSTM) networks is developed to effectively capture long-range dependencies in sensor data. Notably, this study considers both sample class balance and model design as key factors influencing monitoring accuracy and proposes innovative solutions for each, in contrast to previous studies that focused only on one of these aspects. Experimental results on field data show that the proposed DCC-LSTM model achieves a recognition accuracy of 96.38%. When trained on the balanced dataset, six common intelligent models (k-Nearest Neighbor (kNN), decision tree, random forest, LSTM, Gated Recurrent Unit (GRU), and DCC-LSTM) exhibit significant performance improvements, with a maximum accuracy increase of 6.76%. The proposed method can be applied for minor lost circulation recognition in scenarios with limited samples, and is expected to be extended to other drilling anomaly detection tasks or complex downhole condition monitoring in the future.