Scholarlink

<p>Pipelines are among the most cost-effective and widely adopted means of transporting liquids and gases. Multiphase pipelines, which are designed to carry both liquid and gas simultaneously rather than a single fluid, are widely used in modern industrial applications. However, leakage events in such pipelines can lead to severe economic losses, environmental damage, and safety hazards. In addition, leak-related acoustic emission (AE) signals are highly sensitive to feature extraction techniques and are often corrupted by noise in real industrial environments. Although deep learning-based approaches have shown promising performance in single-phase pipeline leak detection, they fail to fully capture both local time-frequency characteristics and global temporal dependencies in multiphase pipeline systems. To address these challenges, this paper proposes an attention-based CNN–LSTM architecture for robust leak detection and size identification using acoustic emission data. Raw acoustic emission signals acquired from industrial fluid pipelines are first preprocessed to reduce noise and operational variability. They are then fed into two parallel branches: a CNN branch that analyzes continuous wavelet transform (CWT)-based two-dimensional representations to extract localized time-frequency features, and an LSTM branch that models the one-dimensional signal to capture global temporal dependencies. The extracted features are subsequently fused to form a vulnerable feature pool, representing a unified feature space that captures leak-sensitive variations in acoustic emission signals, which is further refined using multi-head attention (MHA) to enhance discriminative capability under varying operating conditions. Finally, the attention-enhanced features are classified to distinguish between normal and leak conditions, as well as different leak sizes under varying pressure scenarios. An experimental evaluation using real-world industrial pipeline data demonstrates the effectiveness of the proposed architecture, achieving a classification accuracy of 96.88%, thereby confirming its reliability and suitability for practical multiphase pipeline monitoring applications.</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

Attention mechanism-based CNN-LSTM hybrid deep learning model for industrial pipeline leak detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月19日

查看全文
  • Saqib Umer,
  • Niamat Ullah,
  • Jong-Myon Kim

摘要

Pipelines are among the most cost-effective and widely adopted means of transporting liquids and gases. Multiphase pipelines, which are designed to carry both liquid and gas simultaneously rather than a single fluid, are widely used in modern industrial applications. However, leakage events in such pipelines can lead to severe economic losses, environmental damage, and safety hazards. In addition, leak-related acoustic emission (AE) signals are highly sensitive to feature extraction techniques and are often corrupted by noise in real industrial environments. Although deep learning-based approaches have shown promising performance in single-phase pipeline leak detection, they fail to fully capture both local time-frequency characteristics and global temporal dependencies in multiphase pipeline systems. To address these challenges, this paper proposes an attention-based CNN–LSTM architecture for robust leak detection and size identification using acoustic emission data. Raw acoustic emission signals acquired from industrial fluid pipelines are first preprocessed to reduce noise and operational variability. They are then fed into two parallel branches: a CNN branch that analyzes continuous wavelet transform (CWT)-based two-dimensional representations to extract localized time-frequency features, and an LSTM branch that models the one-dimensional signal to capture global temporal dependencies. The extracted features are subsequently fused to form a vulnerable feature pool, representing a unified feature space that captures leak-sensitive variations in acoustic emission signals, which is further refined using multi-head attention (MHA) to enhance discriminative capability under varying operating conditions. Finally, the attention-enhanced features are classified to distinguish between normal and leak conditions, as well as different leak sizes under varying pressure scenarios. An experimental evaluation using real-world industrial pipeline data demonstrates the effectiveness of the proposed architecture, achieving a classification accuracy of 96.88%, thereby confirming its reliability and suitability for practical multiphase pipeline monitoring applications.