Development of an AI-Driven Digital Twin Framework for Predictive Maintenance of Urban Infrastructure Systems

Abstract

Modern urban infrastructure systems are becoming increasingly vulnerable to deterioration caused by ageing equipment, growing service demands, and environmental stressors, making traditional reactive and preventive maintenance approaches insufficient to ensure long-term system reliability. This study proposes and evaluates a conceptual AI-driven Digital Twin framework for predictive maintenance of urban infrastructure using simulated heterogeneous infrastructure datasets and benchmark validation experiments. An urban infrastructure testbed comprising structural bridge elements, municipal water distribution pipelines, and flexible pavements was used to evaluate the framework. Semantic alignment of multi-source sensing data was performed within an ontology-based Digital Twin environment. Pre-processing involved the application of discrete wavelet denoising and spatio-temporal matrix factorisation, while the pre-processed data were analysed using a Spatio-Temporal Graph Attention Network integrated with Temporal Convolutional Networks. An asymmetric loss function was introduced to prioritise infrastructure safety by imposing greater penalties on false-negative predictions. Framework performance was evaluated through 10-fold rolling-horizon cross-validation and comparison with Vector Autoregression, Support Vector Regression, Long Short-Term Memory, and Graph Convolutional Network models using Mean Absolute Error, Root Mean Squared Error, and the coefficient of determination (R²). The proposed framework achieved the best results for structural asset prediction, with R² = 0.968, MAE = 0.014, and RMSE = 0.021, outperforming all benchmark models. The analysis showed statistically significant differences in favour of the proposed framework (ANOVA: F(4, 45) = 112.43, p < 0.001), while ablation experiments demonstrated the contributions of graph attention, temporal convolution, wavelet denoising, and asymmetric optimisation. Under simulated high-stress conditions, the asymmetric loss function reduced false-negative failure predictions from 4.8% to 0.0% and increased the maintenance warning horizon from 1.1 to 4.2 days before predicted failure. In scalability tests, the Digital Twin architecture demonstrated the potential for near-real-time performance under the experimental conditions, with a cloud inference latency of 64.1 ms for a city-scale infrastructure network of 1,000 monitoring nodes. Overall, integrating semantic Digital Twin and graph-based deep-learning technologies demonstrated the feasibility of the proposed AI-driven predictive maintenance framework.