基於圖神經網路評估震後橋梁連通系統之恢復力

Abstract

面對地震災害時，交通運輸網絡的恢復力對於維持城市中緊急救護功能和減少對關鍵基礎設施的干擾至關重要。橋梁作為交通運輸網路的主要聯通工具，易受到地震影響而發生破壞，進而降低交通網路整體的功能性。本研究提出了一個基於卷積神經網路（convolutional neural network, CNN）和圖神經網絡（graph neural network, GNN）估計橋梁連通系統震後恢復力的方法。此方法利用CNN基於交通即時影像監視器資料評估橋梁功能現況。接著，利用GNN網絡的拓撲結構和連接性來建立道路與道路之間的關係，來預測地震對交通網絡功能之影響。即使對相對較小的資料集進行了訓練（CNN的200張影像資料和1200個震後模擬之圖），這些模型已表現非常不錯，CNN達到了96.5%的分類準確率，而圖轉換模型(graph transformer)的R平方達到0.990。此方法考慮了地震會造成的橋梁損失資訊、交通車流資料和交通網絡拓撲，提供了震後交通路網的恢復力評估，並有效地模擬以最大化增益橋梁恢復力之優先順序。透過台北市城際橋梁網絡的案例研究，展示了本研究提出的方法的效能。

Ensuring the resilience of transportation networks in the face of seismic hazards is vital for rapid recovery and minimizing disruptions to critical infrastructure. Bridges are especially prone to damage from earthquakes, posing significant risks to overall network integrity. This work presents a novel approach to assess the resilience of road-bridge networks by combining a fine-tuned EfficientNet model with a Graph Neural Network (GNN). EfficientNet is utilized to evaluate bridge serviceability based on data from traffic surveillance cameras. Concurrently, GNNs harness the topological structure and interconnectivity of the network to model dependencies and predict the effects of seismic events on network functionality. This comprehensive methodology integrates seismic hazard information, traffic data, and network topology to provide a robust resilience assessment. The combined inputs enables the estimation of network functionality and the development of a bridge restoration prioritization strategy using GNNs. Despite being trained on a relatively small dataset (200 images for EfficientNet and 1200 simulated graphs), the models perform impressively, achieving a 96.5% classification accuracy with EfficientNet and an R-squared score of 0.990 with the Graph Transformer. This approach effectively simulates the prioritization of bridge restorations to maximize functionality gains. The method's efficacy is demonstrated through a case study of the road-bridge network in Taipei City.