Resilience Evaluation of Urban Tunnel Operations Based on Multi-Source Data and Explainable Artificial Intelligence

As critical nodes in urban transportation networks, the operational resilience of urban tunnels directly affects roadway network efficiency and public safety. Existing studies on tunnel resilience exhibit deficiencies in the attention paid to routine incidents, the interpretability of evaluation methods, and the robustness of underlying data. To address these gaps, this paper constructs a three-layer resilience evaluation framework (pre-event, during-event, and post-event) based on 1123 accident records, hourly traffic flow data, and expert scoring data from 10 urban tunnels in Shanghai in 2024. The framework integrates the Analytic Hierarchy Process (AHP) and min–max normalization to derive comprehensive resilience scores. Furthermore, two explainable artificial intelligence (XAI) techniques—SHAP (SHapley Additive exPlanations) feature importance analysis and K-means clustering—are introduced to identify key drivers of resilience differentiation and typical resilience patterns, respectively. Empirical results reveal that (1) half of the 10 tunnels achieve a B-level rating or above, while C/D-level tunnels still account for 50%, indicating considerable room for improvement; (2) SHAP analysis suggests that during-event absorption capacity contributes most significantly to resilience differences (50.24%), with the average traffic reduction rate being the most influential single indicator within this model; and (3) K-means clustering classifies the tunnels into four patterns: high-resilience balanced type, pre-event weak type, post-event weak type, and during-event weak type, each corresponding to distinct resilience enhancement strategies. The methodological framework developed in this paper effectively enables the quantification, interpretation, and pattern recognition of tunnel resilience, providing data-driven decision support for urban tunnel operation and management. It should be noted that the A–D ratings are sensitive to the normalization method: under an alternative normalization, 7 out of 10 tunnels change grade. The ratings should therefore be interpreted as relative to the selected sample rather than as absolute performance levels.