NEXUS SHIELD: A Machine Learning Based Digital Twin Framework for Fault Diagnosis and Safety Constrained Recovery in Water Distribution Networks

Pressure drops in a water distribution network (WDN) are ambiguous. A falling manifold head at a junction might indicate a ruptured branch pipe, an upstream clog, higher consumer demand, or simple transducer drift. Worse, acting directly on a classifier output is dangerous: slamming a solenoid valve shut to isolate a suspected leak triggers a Joukowsky pressure surge that can burst the main line. We present NEXUS SHIELD, an EPANET driven fault diagnosis and control pipeline that separates leak classification from valve actuation. Using 4,000 simulated 60 s operating episodes sampled at 10 Hz across four regimes Nominal operation, Branch A leakage, Branch B clogging, and sensor drift we benchmark an XGBoost classifier against Random Forest, L2 Logistic Regression, and static thresholding. On a 600 episode held out test split, XGBoost classifies network state at 98.17% accuracy and macro F1 with a 0.61% one vs rest false alarm rate. When a rupture is confirmed, the controller rejects instantaneous shutoff which spikes manifold head to 21.6 m against a 16.0 m pipe limit and instead pre opens a bypass line while slewing the isolation valve along a rate limited S curve (3.5°/tick), capping peak transient head at 14.4 m.