Dynamic seismic damage assessment of distributed infrastructure systems using graph neural networks and semi-supervised machine learning

A methodology is presented for performing dynamic seismic damage assessment of distributed infrastructure systems using graph neural networks and semi-supervised machine learning. To achieve this goal, a pre-event damage assessment is performed using either traditional fragility-based models or a machine learning classification algorithm trained on historical damage data. Then, a graph-neural network is implemented to perform semi-supervised learning and update the pre-event predictions as observations of actual damage become available during the post-earthquake inspection process. The methodology is demonstrated on the pipe network for the City of Napa, California water distribution system. A dataset of pipes damaged during the 2014 M 6.0 earthquake is used for validation purposes. A conventional neural network classification model is first trained on a portion of the observed pipe damage and used to perform the pre-event damage assessment i.e., supervised learning. Following the earthquake, a graph neural network model is employed to update the damage estimates given the information incrementally collected during the inspection process. The evaluation results show that the neural network supervised learning model provides better pre-event damage estimates than the existing repair rate-based model. Also, the graph neural network models can provide improved damage quantification given partial information collected during the inspection process.

Automated summary

This study presents a methodology for dynamic seismic damage assessment of distributed infrastructure systems using graph neural networks and semi-supervised machine learning. The approach combines traditional fragility-based models or machine learning classification algorithms with actual damage observations to update pre-event damage assessments. The evaluation results show that the methodology provides better damage estimates than existing models and can quantify damage accurately based on partial information.