thanglexuan/Z24-dataset-processed

---
license: mit
task_categories:
- time-series-forecasting
- tabular-classification
- other
task_ids:
- multi-class-classification
language:
- en
tags:
- structural-health-monitoring
- vibration-data
- bridge-monitoring
- time-series
- accelerometer
- civil-engineering
- damage-detection
- kuleuven
- z24-bridge
pretty_name: Z24 Bridge Structural Health Monitoring Dataset
size_categories:
- 1K<n<10K
source_datasets:
- original
paperswithcode_id: z24-bridge
configs:
- config_name: default
data_files:
- split: n/a
path: "*.npy"
dataset_info:
features:
- name: inputs
dtype:
shape: [27, 6000]
dtype: float64
description: Vibration time-series data from 27 accelerometer sensors
- name: labels
dtype: int64
description: Damage state labels
splits:
- name: full-data
num_bytes: 2500000000
num_examples: 1530
download_size: 2500000000
dataset_size: 2500000000
---

# Z24 Bridge Dataset - KULEUVEN

## Dataset Overview

The Z24 Bridge dataset is a crucial benchmark in Structural Health Monitoring (SHM) research, collected from the Z24 bridge in Switzerland by KULEUVEN (Katholieke Universiteit Leuven). This dataset is widely used in structural health monitoring and damage detection research.

## Z24 Bridge Information

The **Z24 Bridge** was a two-span prestressed concrete bridge built in 1963 in Switzerland with the following specifications:

- **Total Length**: 60m (2 spans × 30m each)
- **Width**: 8.3m
- **Structure**: Prestressed concrete box girder
- **Bridge Type**: Two-span continuous
- **Construction Year**: 1963
- **Data Collection Period**: 1998 (before controlled demolition)

## Dataset Structure

### Original Raw Data
```
Shape: (17, 9, 27, 65530)
```

**Dimension meanings:**
- **17**: Number of scenarios/test conditions
- **9**: Number of measurement setups/configurations
- **27**: Number of sensors/measurement points
- **65530**: Length of time points in each measurement

### After Data Cleaning
```
Shape: (17, 9, 27, 60000)
```

**Processing steps:**
- Removed 5530 corrupted or missing data points
- Normalized time series length to 60000 samples
- Maintained original sensor and setup layout

### Final Processed Data (Reshaped)
```
Shape: (1530, 27, 6000)
```

**Reshaping process:**
- **1530 = 17 × 9 × 10**: Total measurement segments
- 17 scenarios
- 9 setups
- 10 segments divided from each long time series
- **27**: Number of sensors (unchanged)
- **6000 = 60000 ÷ 10**: Each segment contains 6000 time samples

## Technical Specifications

### Measurement Parameters
- **Sensor type**: Accelerometers
- **Measurement**: Vibration response data
- **Units**: m/s² (acceleration)

### Data Collection Conditions
- **Environmental conditions**: Controlled and natural conditions
- **Excitation**: Ambient vibration, controlled loading
- **Damage scenarios**: Progressive damage artificially introduced
- **Reference state**: Undamaged baseline measurements

## File Structure

```
Data_Z24_processed/
├── README.md # This documentation file (English)
├── README_vi.md # Vietnamese version
├── inputs.npy # Input data (1530, 27, 6000)
├── labels.npy # Labels for damage states
├── data_visualization.ipynb # Jupyter notebook for data exploration
├── pixi.toml # Environment configuration
└── pixi.lock # Locked dependencies
```

### File Descriptions

#### `inputs.npy`
- **Shape**: (1530, 27, 6000)
- **Type**: numpy.float64
- **Content**: Preprocessed vibration time-series data
- **Sensors (27)**: Accelerometers placed at various locations on the bridge

#### `labels.npy`
- **Shape**: Corresponding to sample count (1530,)
- **Type**: Depends on classification task
- **Content**:
- Damage state labels
- Environmental condition indicators
- Or regression targets

## Research Applications

This dataset is commonly used for:

### 1. **Damage Detection**
- Binary classification: damaged vs undamaged
- Multi-class: different damage severities
- Anomaly detection algorithms

### 2. **Damage Localization**
- Spatial localization of damage
- Sensor fusion techniques
- Modal analysis approaches

### 3. **Machine Learning Applications**
- Deep learning for SHM
- Feature extraction methods
- Time series classification
- Transfer learning studies

### 4. **Signal Processing Research**
- Modal parameter identification
- Frequency domain analysis
- Time-frequency analysis
- Noise robustness studies

## Dataset Characteristics

### Advantages
- ✅ **Realistic data**: Collected from real-world structure
- ✅ **Controlled damage**: Progressive damage scenarios
- ✅ **Multiple sensors**: Rich spatial information
- ✅ **Benchmark status**: Widely used for comparison
- ✅ **Well-documented**: Extensive literature available

### Challenges
- ⚠️ **Environmental effects**: Temperature, humidity variations
- ⚠️ **Operational variability**: Traffic, loading conditions
- ⚠️ **Class imbalance**: Limited damage data vs normal conditions
- ⚠️ **Noise levels**: Real-world measurement noise
- ⚠️ **Sensor placement**: Fixed sensor configuration

## Data Processing Pipeline

```python
# Data processing flow from raw to final
Raw Data (17, 9, 27, 65530)
↓ [scenarios, setups, sensors, timepoints]
[Data Cleaning & Quality Control]
↓
Cleaned Data (17, 9, 27, 60000)
↓
[Segmentation & Reshaping]
↓
Final Data (1530, 27, 6000)
↓ [segments, sensors, timepoints_per_segment]
```

### Preprocessing Steps Performed
1. **Quality Control**: Removed corrupted measurements
2. **Length Normalization**: Trimmed/padded to 60000 samples
3. **Segmentation**: Divided into 6000-sample segments
4. **Reshaping**: Flattened scenario × setup × segment dimensions

## Usage Instructions

### Loading Data
```python
import numpy as np

# Load input data and labels
inputs = np.load('inputs.npy') # Shape: (1530, 27, 6000)
labels = np.load('labels.npy') # Shape: (1530,)

print(f"Input shape: {inputs.shape}")
print(f"Labels shape: {labels.shape}")
```

### Basic Analysis
```python
# Statistical analysis
print(f"Input range: [{inputs.min():.4f}, {inputs.max():.4f}]")
print(f"Input mean: {inputs.mean():.4f}")
print(f"Input std: {inputs.std():.4f}")

# Label distribution
unique_labels, counts = np.unique(labels, return_counts=True)
print(f"Unique labels: {unique_labels}")
print(f"Label counts: {counts}")
```

### Visualization
Use the provided `data_visualization.ipynb` notebook to visualize and explore the dataset.

## References

### Key Papers
1. **Worden, K., & Manson, G.** (2007). The application of machine learning to structural health monitoring. *Philosophical Transactions of the Royal Society A*, 365(1851), 515-537.

2. **Reynders, E., et al.** (2008). Output-only structural health monitoring in changing environmental conditions by means of nonlinear system identification. *Structural Health Monitoring*, 7(4), 243-268.

3. **Maeck, J., & De Roeck, G.** (2003). Dynamic bending and torsion stiffness derivation from modal curvatures and torsion rates. *Journal of Sound and Vibration*, 265(1), 153-170.

### Dataset References
- **KULEUVEN SHM Database**: Structural Health Monitoring research group
- **Z24 Bridge Benchmark**: Community standard in SHM research
- **Original Studies**: Search literature with keywords "Z24 bridge" + "KULEUVEN"

## Contact & Support

- **Institution**: KULEUVEN (Katholieke Universiteit Leuven)
- **Department**: Civil Engineering
- **Research Group**: Structural Mechanics
- **Website**: [KULEUVEN Civil Engineering](https://www.kuleuven.be/)

## License & Citation

When using this dataset in research, please cite the original papers and acknowledge KULEUVEN university.

```bibtex
@misc{z24_bridge_dataset,
title={Z24 Bridge Structural Health Monitoring Dataset},
author={KULEUVEN},
institution={Katholieke Universiteit Leuven},
year={1998},
note={Processed version: 17 scenarios, 9 setups, 27 sensors, reshaped to (1530, 27, 6000)}
}
```

## Data Structure Summary

| Dimension | Original | Cleaned | Final | Description |
|-----------|----------|---------|-------|-------------|
| Scenarios | 17 | 17 | 17×10=170* | Test conditions |
| Setups | 9 | 9 | 9×10=90* | Measurement configurations |
| Sensors | 27 | 27 | 27 | Accelerometer locations |
| Timepoints | 65530 | 60000 | 6000 | Samples per segment |
| **Total Shape** | **(17,9,27,65530)** | **(17,9,27,60000)** | **(1530,27,6000)** | **Final format** |

*\*Each original time series was divided into 10 segments, hence 1530 = 17×9×10*

---

*Dataset processed and prepared for Structural Health Monitoring and Machine Learning research.*