Water Pipeline Leakage Classification
A comparative study of ML and deep learning methods for urban water pipeline leak detection.
Project Snapshot
Conducted at UNIST AI Graduate School in collaboration with ETRI as part of a government-funded research project on urban water pipeline monitoring. Benchmarked classical ML and deep learning models on real-world frequency-domain sensor data, culminating in a Graph Transformer achieving 96.4% accuracy with perfect micro-leak detection.
Problem
Urban water pipeline networks lose significant volumes of water through undetected leaks. Traditional inspection methods are costly and slow. The challenge was to classify pipeline sensor readings into three classes — normal, general leak, and micro leak — using frequency-domain time-series data from 391 sensors in Daegu, South Korea.
Approach
- Processed real-world data: 512 frequency signals per day over 5 consecutive days from 391 water-pipeline sensors
- Benchmarked KNN, Random Forest, CNN, fine-tuned MLP, and a custom Graph Transformer
- Designed the Graph Transformer to leverage attention mechanisms across distributed sensor signals for spatial pattern recognition
- Performed hyperparameter tuning and class-wise evaluation for robust three-class classification
Tech Stack
Python, PyTorch, Scikit-learn, NumPy, Pandas, Matplotlib
Results
- Graph Transformer achieved overall accuracy of ≈ 96.4%
- Perfect precision and recall (1.000) for micro-leak detection
- Improved class-wise robustness and interpretability over classical baselines
- Published at Korea Software Congress (December 2025)