MLForgeStructureDamageDetection

🏗️ Structural Health Monitoring System

A machine learning-based system for detecting structural damage in a three-story aluminum building structure using vibration data.
This project was developed for the ML Forge competition as part of Shilp '25.

---

📚 Table of Contents

• Project Overview
• Dataset Description
• Installation
• Usage Guide
• Files and Directories
• Results and Visualization
• Troubleshooting
• Future Improvements
• License
• Acknowledgments

---

📌 Project Overview

This system helps engineers detect structural damage by analyzing vibration data from a three-story aluminum structure.

It distinguishes between:

• ✅ Normal environmental variations (e.g., temperature changes, aging)
• ❌ Actual structural damage

🏗️ Test Environment Simulation

• Nonlinear damage: Introduced via suspended column and adjustable bumper
• Environmental effects: Simulated using added mass or reduced stiffness

💡 Why Our Solution Stands Out

• Extracts meaningful patterns from complex vibration data
• Achieves 97% accuracy using a Random Forest model
• Easy to analyze new structural data
• Includes intuitive visualizations for clear insights

---

📊 Dataset Description

The dataset includes vibration readings from a specially built aluminum structure.

🔧 What’s Measured?

• Force Transducer: Measures input force
• Four Accelerometers: Measure response on each floor

🏷️ Conditions Simulated

• Normal Baseline:

  • Reference state (e.g., state#13)

• Environmental Changes (Undamaged):

  • Added weight
  • Reduced stiffness

• Damage Conditions:

  • Varying bumper-column gaps for non-linear behavior

• Mixed Conditions:

  • Combine environmental + damage to test model robustness

Each condition contains multiple files with full sensor readings.

---

⚙️ Installation

📥 Clone the Repository

git clone https://github.com/your-username/structural-health-monitoring.git
cd structural-health-monitoring

🐍 Set Up a Virtual Environment

On Windows:

python -m venv venv
venv\Scripts\activate

On macOS/Linux:

python -m venv venv
source venv/bin/activate

📦 Install Required Packages

pip install numpy pandas matplotlib scikit-learn scipy joblib seaborn tqdm

Or use a requirements.txt:

pip install -r requirements.txt

---

🚀 Usage Guide

🧠 Train the Model

python structural_health_monitoring.py

What it does:

• Loads and processes vibration data
• Extracts features
• Trains a Random Forest model
• Evaluates model and generates visualizations
• Saves trained model & processed data

You'll see:

Dataset created with 169 samples and 86 features
Test accuracy: 0.9706
Confusion Matrix:
[[18  0]
 [ 1 15]]

---

🔎 Make Predictions

📂 Analyze All Data Automatically

python predict.py --auto --model structural_damage_model.pkl

🎯 Analyze a Specific Folder

python predict.py --dir "path/to/data/directory" --model structural_damage_model.pkl

📄 Analyze a Single File (with visualization)

python predict.py --file "path/to/data/file.txt" --model structural_damage_model.pkl --visualize

🆘 Get Help

python predict.py --help

---

📈 Prediction Output Example

===== Prediction Summary =====
Total files processed: 169
Overall damage detection: 50 damaged, 119 undamaged
Results saved to prediction_results.csv

---

🔬 Feature Analysis

🧠 What is the model "thinking"?

🔥 Most Important Features

python analyze_features.py --importance --model structural_damage_model.pkl --dataset processed_dataset.csv

🧩 Feature Clustering (PCA/TSNE)

python analyze_features.py --clustering --dataset processed_dataset.csv

🔗 Feature Correlations

python analyze_features.py --correlations --dataset processed_dataset.csv

🚦 Damaged vs. Undamaged Separability

python analyze_features.py --separability --dataset processed_dataset.csv

---

📁 Files and Directories

File	Description
structural_health_monitoring.py	Core pipeline for training and processing
predict.py	Main script to analyze new data
analyze_features.py	Explains model behavior and feature interactions
debug_data_loading.py	Helps troubleshoot dataset issues
structural_damage_model.pkl	Trained model file
processed_dataset.csv	Feature-rich, clean data
feature_importance.png	Visualization of important features
confusion_matrix.png	Model performance summary
requirements.txt	Python dependency list

---

📊 Results and Visualization

✅ Accuracy: 97.06%

• 100% precision for undamaged states
• 93.8% precision for damaged states
• Only 1 misclassification in 34 test samples

🖼️ Visual Insights:

• Feature Importance Plots
• Confusion Matrix
• PCA & t-SNE Embeddings
• Correlation Heatmaps

---

🛠️ Troubleshooting

Issue	Fix
File not found	Check paths and extensions (.txt)
Model loading issues	Ensure model file exists and matches scikit-learn version
Memory errors	Try smaller file batches
Unknown errors	Run debug_data_loading.py

---

🚧 Future Improvements

• Implement deep learning (CNNs, LSTMs)
• Real-time health monitoring
• Web dashboard for live visualization
• Damage localization (multi-class classification)
• Adaptation for different structure types

---

📄 License

This project is under the MIT License. See the LICENSE file for more details.

---

🙏 Acknowledgments

Special thanks to:

• Shilp '25 / ML Forge Team – for organizing the challenge
• Open-source contributors – for the amazing tools and libraries