Quantifying extreme failure scenarios in transportation systems with graph learning.

Area of Science:

• Engineering
• Computer Science
• Statistics

Background:

• Assessing extreme events in complex engineering systems is crucial for reliability and resilience.
• Estimating failure probabilities is computationally expensive due to event rarity and system complexity.
• Traditional importance sampling methods face challenges with high costs in large-scale systems.

Purpose of the Study:

• To propose a novel graph learning approach for efficient statistical analysis of extreme events.
• To reduce the computational burden associated with estimating rare event probabilities in complex networks.
• To enhance the accuracy and efficiency of reliability and resilience assessments.

Main Methods:

• Introduced a graph learning approach: importance sampling based on graph autoencoder (GAE-IS).
• Integrated a modified graph autoencoder (criticality assessor) with cross-entropy-based importance sampling.
• Decoupled component criticality from vulnerability in the analysis workflow.

Main Results:

• GAE-IS demonstrated notable transferability across different network types.
• Achieved significant reductions in computational costs for importance sampling (one to two orders of magnitude).
• Provided more accurate probability estimations for extreme failures in large-scale networks.

Conclusions:

• GAE-IS offers a computationally efficient and accurate method for extreme event probability estimation.
• The approach is highly effective for large-scale engineering systems, particularly road networks.
• This methodology advances system design, reliability, and resilience assessment.