Scholarlink

In the field of Structural Health Monitoring (SHM), the application of deep learning models for analyzing time-series data has garnered significant attention. One-dimensional convolutional neural networks (1DCNN) are commonly used but face limitations in effectively handling long datasets. Therefore, this study proposes a novel approach by combining 1DCNN with the Squeeze-and-Excitation (SE) mechanism (SE-1DCNN) and Long Short-Term Memory (LSTM) networks to accurately classify structural damage. This combination leverages the spatial feature extraction and attention mechanism of SE-1DCNN alongside LSTM’s capability to process long-term time-series data. The model is trained and evaluated using an experimental dataset collected from a steel frame structure instrumented with multiple accelerometers under various damage scenarios. The proposed SE-1DCNN-LSTM model achieves an accuracy of 96.7% on the training set and 95.3% on the test set, outperforming the traditional 1DCNN-LSTM model. These results confirm that integrating SE-1DCNN and LSTM enhances damage classification accuracy and demonstrates strong potential for real-world SHM applications.

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 Classification of Steel Frames Using Long Short-Term Memory and Fully Convolutional Network Models

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Truong Thanh Chung,
  • Tran Tien Son,
  • Le Van Vu,
  • Luong Nguyen-Duc,
  • Tran Ngoc Hoa

摘要

In the field of Structural Health Monitoring (SHM), the application of deep learning models for analyzing time-series data has garnered significant attention. One-dimensional convolutional neural networks (1DCNN) are commonly used but face limitations in effectively handling long datasets. Therefore, this study proposes a novel approach by combining 1DCNN with the Squeeze-and-Excitation (SE) mechanism (SE-1DCNN) and Long Short-Term Memory (LSTM) networks to accurately classify structural damage. This combination leverages the spatial feature extraction and attention mechanism of SE-1DCNN alongside LSTM’s capability to process long-term time-series data. The model is trained and evaluated using an experimental dataset collected from a steel frame structure instrumented with multiple accelerometers under various damage scenarios. The proposed SE-1DCNN-LSTM model achieves an accuracy of 96.7% on the training set and 95.3% on the test set, outperforming the traditional 1DCNN-LSTM model. These results confirm that integrating SE-1DCNN and LSTM enhances damage classification accuracy and demonstrates strong potential for real-world SHM applications.