Scholarlink

<p>Addressing the challenge of identifying gas sources within underground gas networks where biogas interference complicates detection, this study proposes a machine learning-based gas source classification method tailored for valve well scenarios. This approach relies solely on CH₄ concentration time series data. Through theoretical analysis of the generation mechanisms and diffusion processes of gas leaks versus biogas accumulation, a novel three-tier hierarchical nested multi-scale time window (7&#xa0;days/3&#xa0;days/1&#xa0;day) feature extraction and fusion method is introduced. Utilising an independently developed monitoring platform, continuous observations spanning 20&#xa0;months were conducted across 1,351 gas valve chambers, establishing a feature vector set with spatio-temporal resolution. Performance differences among three classifiers—Random Forest (RF), XGBoost, and KNN—were systematically evaluated using precision-recall curves and ROC curves. Experiments demonstrated that the XGBoost model achieved superior recognition accuracy at 68.63%, outperforming both Random Forest (67.67%) and KNN (65.38%). By optimising feature weighting strategies, this approach enables online identification and early warning of leakage incidents, providing an efficient technical pathway for underground pipeline network safety monitoring. Experimental validation confirms the framework's engineering feasibility in limited deployment scenarios, though comprehensive implementation requires further field-based empirical research.</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

Study on leakage identification of buried gas pipelines under biogas disturbances

  • 期刊论文

  • 发表日期：

    2026年4月9日

查看全文
  • Jun Li,
  • Xu Luchen,
  • Shi Jiakang,
  • Li Meng,
  • Jin Zengfeng

摘要

Addressing the challenge of identifying gas sources within underground gas networks where biogas interference complicates detection, this study proposes a machine learning-based gas source classification method tailored for valve well scenarios. This approach relies solely on CH₄ concentration time series data. Through theoretical analysis of the generation mechanisms and diffusion processes of gas leaks versus biogas accumulation, a novel three-tier hierarchical nested multi-scale time window (7 days/3 days/1 day) feature extraction and fusion method is introduced. Utilising an independently developed monitoring platform, continuous observations spanning 20 months were conducted across 1,351 gas valve chambers, establishing a feature vector set with spatio-temporal resolution. Performance differences among three classifiers—Random Forest (RF), XGBoost, and KNN—were systematically evaluated using precision-recall curves and ROC curves. Experiments demonstrated that the XGBoost model achieved superior recognition accuracy at 68.63%, outperforming both Random Forest (67.67%) and KNN (65.38%). By optimising feature weighting strategies, this approach enables online identification and early warning of leakage incidents, providing an efficient technical pathway for underground pipeline network safety monitoring. Experimental validation confirms the framework's engineering feasibility in limited deployment scenarios, though comprehensive implementation requires further field-based empirical research.