An intelligent multi-fluid machine learning framework for predicting leaks in generalized pipeline systems

Pipeline leakage poses significant challenges in municipal, industrial, chemical, and fuel transportation networks. Conventional leak prediction approaches, predominantly designed for water, often exhibit limited accuracy when applied to fluids with varying densities and viscosities. This study proposes a physics-informed machine learning framework combining a simplified laminar hydraulic model, density-dependent analytical leak correlations, and supervised learning techniques to predict leakage size and location from hydraulic measurements. Four fluids, namely water (1000 kg/m³), ethanol (800 kg/m³), diesel (850 kg/m³), and glycerin (1200 kg/m³), were examined using 1000 simulated scenarios for each fluid. Pipe diameter and inlet velocity were selected to maintain laminar flow conditions (Re ≤ 2000). The hydraulic framework incorporated continuity, Hagen–Poiseuille flow, and orifice discharge equations, with validation against established analytical solutions and reported discharge coefficients. Results demonstrated that leakage size primarily depends on flow deficit, whereas leak jet velocity remains unaffected by leak dimensions. Five predictive models—Ridge Regression, Random Forest, Gradient Boosting, Support Vector Regression, and Multi-Layer Perceptron—were evaluated using an 80:20 training-testing split, five-fold cross-validation, grid-search hyperparameter tuning, fixed random seeds, and five independent runs. Pressure was exclusively considered an input variable. Gradient Boosting demonstrated superior performance in estimating leak size (MAE ≈ 0.25 mm; R² ≈ 0.977) and location (MAE ≈ 1.7 m; R² ≈ 0.92). Feature importance findings aligned with fundamental hydraulic principles, while cross-fluid evaluations confirmed the ensemble models' reliability and generalizability across different fluid properties.