Scholarlink

<p>The recent increase in safety incidents in subway tunnels, caused by complex geological conditions and dynamic vehicle loads, highlights the urgent need for structural health monitoring. Internal structural damage is hard to detect through manual inspection; however, changes in the structural vibration response can effectively reveal hidden defects, offering a significant breakthrough for intelligent diagnosis. While vibration data-driven deep learning damage detection has gained attention, existing methods often struggle to integrate temporal dependencies and spatial features in complex scenarios effectively. Specific limitations exist in identifying multiple damages within confined tunnel spaces, and manual adjustments to network parameters are often required based on human experience. This reduces efficiency and compromises objectivity. Therefore, this study presents a novel global damage identification method for shield tunnel sections, using KOA-CNN-BiLSTM to analyze structural vibration signals. The method utilizes Bidirectional Long Short-Term Memory (BiLSTM) networks to improve the feature extraction capabilities of convolutional neural networks (CNN), capturing features by simultaneously learning dependencies in time-series data from both directions. Combined with the Kepler Optimization Algorithm (KOA), it optimizes the learning rate, convolutional kernel size, and the number of hidden neurons, significantly reducing training costs. When applied to tunnel section structural damage identification, this method extracts the optimal damage information from raw acceleration signals. It facilitates comprehensive damage identification and the automatic localization of damage-sensitive features, thereby enhancing the objectivity of the assessment. Verified through large-scale tunnel-soil model experiments, this method achieved an accuracy rate of 95.19% in identifying and locating damages of various degrees. Its performance significantly surpasses that of a single CNN model in complex scenarios, such as when multiple damages exist within the same tunnel ring. As a pioneering application of deep learning in structural damage identification of subway tunnel linings, this method holds great potential for practical engineering applications and further 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

Damage identification of subway shield tunnel segment structure based on KOA-CNN-BiLSTM

  • 期刊论文

  • 发表日期：

    2025年10月22日

查看全文
  • Tao Li,
  • Zhongyu Zhang,
  • Kangkang Zheng,
  • Rui Hou,
  • Dongwei Ren,
  • Bo Liu,
  • Qian Chen

摘要

The recent increase in safety incidents in subway tunnels, caused by complex geological conditions and dynamic vehicle loads, highlights the urgent need for structural health monitoring. Internal structural damage is hard to detect through manual inspection; however, changes in the structural vibration response can effectively reveal hidden defects, offering a significant breakthrough for intelligent diagnosis. While vibration data-driven deep learning damage detection has gained attention, existing methods often struggle to integrate temporal dependencies and spatial features in complex scenarios effectively. Specific limitations exist in identifying multiple damages within confined tunnel spaces, and manual adjustments to network parameters are often required based on human experience. This reduces efficiency and compromises objectivity. Therefore, this study presents a novel global damage identification method for shield tunnel sections, using KOA-CNN-BiLSTM to analyze structural vibration signals. The method utilizes Bidirectional Long Short-Term Memory (BiLSTM) networks to improve the feature extraction capabilities of convolutional neural networks (CNN), capturing features by simultaneously learning dependencies in time-series data from both directions. Combined with the Kepler Optimization Algorithm (KOA), it optimizes the learning rate, convolutional kernel size, and the number of hidden neurons, significantly reducing training costs. When applied to tunnel section structural damage identification, this method extracts the optimal damage information from raw acceleration signals. It facilitates comprehensive damage identification and the automatic localization of damage-sensitive features, thereby enhancing the objectivity of the assessment. Verified through large-scale tunnel-soil model experiments, this method achieved an accuracy rate of 95.19% in identifying and locating damages of various degrees. Its performance significantly surpasses that of a single CNN model in complex scenarios, such as when multiple damages exist within the same tunnel ring. As a pioneering application of deep learning in structural damage identification of subway tunnel linings, this method holds great potential for practical engineering applications and further research.