Anomaly Detection in Urban Utilities via Multimodal IoT-GIS Big Data Analytics

Abstract

Urban utility networks - potable water distribution systems and low-voltage electrical feeders - form the physiological backbone of contemporary smart cities, yet their fault-detection pipelines remain largely siloed, relying on isolated sensor thresholds that disregard the spatial topology through which faults actually propagate. This paper proposes a multimodal Internet of Things-Geographic Information System (IoT-GIS) big data analytics framework that fuses heterogeneous sensor telemetry - pressure, flow, voltage, current, and transformer temperature - with spatial network graphs derived from pipeline and feeder topology to detect anomalies such as leakage, pipe burst, pressure drop, voltage fluctuation, power theft, transformer failure, and sensor fault. The architecture couples edge gateways, a distributed big data platform, and a GIS-indexed spatial graph database with a hybrid Graph Neural Network-Long Short-Term Memory (GNN-LSTM) model that jointly learns spatial correlation among topologically adjacent assets and temporal dynamics within individual sensor streams. Experiments on EPANET-simulated water networks and smart-meter electrical datasets, benchmarked against Random Forest, XGBoost, CNN, LSTM, and a standalone GNN, show that the proposed hybrid model attains 96.8 percent accuracy and a 0.978 ROC-AUC, outperforming the strongest baseline by 4.6 percentage points while reducing false alarms across six anomaly categories. The results confirm that fusing spatial-graph topology with temporal telemetry yields more robust, interpretable, and scalable fault localization than modality-isolated approaches, offering a deployable blueprint for real-time smart-city utility monitoring..