Proactive maintenance of urban transport infrastructure is critical for ensuring public safety and operational continuity within the evolving landscape of modern smart cities. However, traditional strategies often rely on periodic manual inspections, which are frequently inefficient, costly, and fail to detect early-stage structural degradation. This paper presents an innovative framework for predictive maintenance that integrates high-fidelity Digital Twin (DT) technology with advanced Artificial Intelligence (AI) algorithms, specifically Long Short-Term Memory (LSTM) networks. The proposed system establishes a robust virtual replica of physical transport assets, continuously synchronized with real-time data from dense IoT sensor networks using Kalman filtering for precise state estimation. The methodology details a multi-domain DT development process that simulates structural behavior under diverse loading and environmental conditions. Advanced signal processing techniques, including Root Mean Square (RMS) and Fast Fourier Transform (FFT), are employed to extract salient features from raw sensor data. These features feed into trained LSTM networks designed to forecast structural defects and remaining useful life. Experimental validation on a 1:50 scaled bridge model-subjected to simulated fatigue, corrosion, and settlement-demonstrates a significant improvement in defect prediction accuracy, exceeding 95% with an average lead time of 30 days. This approach transforms maintenance from reactive to predictive, effectively enhancing safety, extending asset lifespan, and optimizing management budgets for critical urban transport systems.