Exploring the role of explainable AI in urban rail transit operations and passenger services

Interpretable artificial intelligence (AI) is essential for addressing the increasing complexity of urban railway systems, were dynamic passenger flows and aging infrastructure result in frequent cascading failures. This study introduces a topology-sensitive Graph Neural Network Explainability (GNN-XAI) framework tailored for fault diagnosis and risk management in urban railways. The framework integrates topology-aware explanations, real-time risk visualization, and engineering ontology alignment to bridge the gap between predictive accuracy and interpretability. Node criticality is evaluated using structural influence and operational data, enabling precise identification of failure-origin nodes. Real-time cascading risk visualization achieves sub-second latency, allowing operators to monitor evolving risks dynamically. Additionally, a Virtual Fault Injection Platform (VFIP) simulates diverse failure scenarios, providing adaptive training environments that enhance diagnostic accuracy, explanation coherence, and response efficiency. Empirical validation on synthetic networks demonstrated resilience against adversarial disruptions, while benchmarking with the Tokyo Metro dataset showed superior performance, with an 11–19% improvement in F1-scores compared to baseline models. The GNN-XAI framework not only ensures robustness and computational efficiency but also extends its impact to engineering education, preparing trainees for real-world challenges in safety-critical systems. This study establishes a foundational approach for deploying interpretable AI in urban railway operations, promoting reliable and actionable decision-making in increasingly complex infrastructure environments.