ResUNet4T: Deep Learning for Structural Damage Detection Using Time-Series Data

---

📄 Paper

ResUNet4T: A potential deep learning model for damage detection based on a numerical case study of a large-scale bridge using time-series data
Engineering Structures, 2025, 120668
🔗 https://doi.org/10.1016/j.engstruct.2025.120668

---

🏗️ Overview

ResUNet4T is a novel deep learning architecture that combines the U-Net encoder–decoder structure with Residual Blocks for one-dimensional time-series classification — specifically designed for structural damage detection in large-scale civil engineering structures.

The model is validated through a comprehensive numerical case study of a large-scale bridge, demonstrating its effectiveness in identifying damage locations and severities from vibration time-series responses.

Key Contributions

• Proposed a 1D Residual U-Net (ResUNet4T) model tailored for time-series-based structural health monitoring (SHM).
• Integrated Layer Normalization and residual shortcut connections to improve gradient flow and generalization.
• Systematically compared against three baseline models: 1DCNN, LSTM, and DCNN-LSTM.
• Demonstrated high classification accuracy across 16 damage scenarios on a large-scale bridge numerical model.

---

🧠 Model Architecture

Input (samples × time steps)
        │
  ┌─────▼─────┐
  │  Encoder  │   Residual Block (16 filters) → MaxPooling1D
  │           │   Residual Block (32 filters) → MaxPooling1D
  └─────┬─────┘
        │
  ┌─────▼─────┐
  │   Bridge  │   Residual Block (64 filters)
  └─────┬─────┘
        │
  ┌─────▼─────┐
  │  Decoder  │   UpSampling1D + Skip Connection → Residual Block (32 filters)
  │           │   UpSampling1D + Skip Connection → Residual Block (16 filters)
  └─────┬─────┘
        │
  GlobalAveragePooling1D
        │
  Dense(num_classes, softmax)
        │
     Output

Each Residual Block consists of:

• Two Conv1D layers with LayerNormalization
• A 1×1 shortcut convolution to match dimensions
• Element-wise Add followed by ReLU activation

---

📂 Repository Structure

Code/
├── ResUNet.py                  # Main model definition and training script
├── requirement.txt             # Python dependencies
├── Result.xlsx                 # Summary of benchmark results
│
├── Data/
│   ├── input_data_v1.npy       # Input dataset (preprocessed, full)
│   └── output_data_v1.npy      # Labels (preprocessed)
│
├── Notebooks/
│   ├── Data_processing.ipynb   # Data preprocessing pipeline
│   ├── ResUNet.ipynb           # Model training & evaluation (Linux)
│   └── ResUNet_macos.ipynb     # Model training & evaluation (macOS)
│
├── Model/
│   ├── model_ResUNet_*.h5      # Saved ResUNet4T model weights
│   ├── DCNN_model_*.h5         # Saved 1DCNN model weights
│   ├── LSTM_model_*.h5         # Saved LSTM model weights
│   ├── DCNN_LSTM_model_*.h5    # Saved DCNN-LSTM model weights
│   └── history_*.json          # Training history logs
│
└── Figure/
    └── ...                     # Generated plots and visualizations

---

⚙️ Installation

Prerequisites

• Python ≥ 3.8
• TensorFlow ≥ 2.x (with GPU support recommended)

Data Split

Split	Ratio
Training	70%
Validation	15%
Test	15%

---

Benchmark Models Compared

Model	Description
ResUNet4T ⭐	Proposed: 1D Residual U-Net
1DCNN	1D Convolutional Neural Network
LSTM	Long Short-Term Memory Network
DCNN-LSTM	Hybrid CNN + LSTM

---

📋 Dependencies

Package	Purpose
tensorflow	Deep learning framework
numpy	Numerical computing
pandas	Data handling & result reporting
scikit-learn	Train/test split, metrics
matplotlib	Visualization
seaborn	Enhanced plotting
tqdm	Progress bars

---

📖 Citation

If you use this code or find this work helpful, please cite:

@article{le2025resunet4t,
  title={ResUNet4T: A potential deep learning model for damage detection based on a numerical case study of a large-scale bridge using time-series data},
  author={Le-Xuan, Thang and Nguyen-Chi, Thanh and Bui-Tien, Thanh and Tran-Ngoc, Hoa},
  journal={Engineering Structures},
  volume={340},
  pages={120668},
  year={2025},
  publisher={Elsevier}
}

---