Scholarlink

<p>In oilfield power operations, accurately recognizing workers’ pole-climbing behavior is critical for operational safety. The task is challenged by subtle fine-grained action differences, ambiguous state transitions, and long-term temporal dependencies, while conventional methods suffer from slow response and inaccurate boundary detection. To address these issues, we propose a dual-branch temporal recognition network based on boundary-aware and adaptive response mechanisms. Built upon R(2 + 1)D and ActionFormer, the network integrates a long-short temporal fusion structure to capture both global context and local variations. A boundary-aware attention module focuses on key transition frames, and a multi-objective loss function enhances temporal consistency and prediction smoothness. An adaptive response strategy dynamically adjusts judgment pace according to confidence levels. Extensive experiments on a self-built Oilfield Power Pole Climbing Dataset (OPPCD) show that the proposed method achieves 97.57% Top-1 accuracy and 97.58% F1-score, significantly outperforming mainstream baselines in Top-1 accuracy, boundary F1 score, and mean Average Precision (mAP).</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

Oilfield power operation behavior recognition based on a boundary-aware and adaptive mechanism in a dual-branch temporal network

  • 期刊论文

  • 发表日期：

    2026年5月25日

查看全文
  • Huizhen Zhang,
  • Guohao Qin,
  • Yuhan Huo,
  • Yanqing Sun,
  • Weijian Ren,
  • Siyu Chen

摘要

In oilfield power operations, accurately recognizing workers’ pole-climbing behavior is critical for operational safety. The task is challenged by subtle fine-grained action differences, ambiguous state transitions, and long-term temporal dependencies, while conventional methods suffer from slow response and inaccurate boundary detection. To address these issues, we propose a dual-branch temporal recognition network based on boundary-aware and adaptive response mechanisms. Built upon R(2 + 1)D and ActionFormer, the network integrates a long-short temporal fusion structure to capture both global context and local variations. A boundary-aware attention module focuses on key transition frames, and a multi-objective loss function enhances temporal consistency and prediction smoothness. An adaptive response strategy dynamically adjusts judgment pace according to confidence levels. Extensive experiments on a self-built Oilfield Power Pole Climbing Dataset (OPPCD) show that the proposed method achieves 97.57% Top-1 accuracy and 97.58% F1-score, significantly outperforming mainstream baselines in Top-1 accuracy, boundary F1 score, and mean Average Precision (mAP).