Leak Detection Physics-Informed Synthetic Dataset

Associated Publication: Cross-Dataset Generalization for Water Leak Detection: A Three-Dataset Transfer Learning Benchmark (DOI pending)

Overview:2,000-record synthetic dataset for cross-domain transfer learning evaluation in water distribution networks. Complements two Kaggle datasets (DS1, DS2) to enable first-ever 3×3 transfer matrix benchmarking.

Dataset Characteristics:

Records: 2,000

Leak Prevalence: 13.8% (276 positives)

Features: Pressure(PSI), FlowRate(gal/min), Temperature(F°), PipeAge(years), PipeDiameter(mm), SensorLocationID

Leak Generation: Physics-informed logistic function:P(leak) = sigmoid(-pressure + flow_deviation + age + (1/diameter) + noise)

Key Design Choices:

Unit mismatch: PSI/gal-min/F° (vs. bar/L-min/C° in DS1/DS2) tests scale robustness

Rich positives: 276 events (vs. DS2's 19) enables stable target-only baselines

Controlled complexity: Parametric physics model ensures separability while testing transfer

Usage:

Within-dataset baseline: ROC-AUC 0.972-0.980

Cross-dataset transfer: DS1↔DS3 bidirectional 0.946

Full analysis code: Included (reproduce Tables 2-5)

License: CC0 1.0 (public domain dedication)

Keywords: water leak detection, transfer learning, domain adaptation, tabular data, synthetic benchmark, cold-start learning, sensor data