Spatially aware Markov chain-based deterioration prediction of bridge components using a Graph Transformer

Abstract

This study proposes a Markov chain-based deterioration prediction framework that incorporates spatial relationships between structural components. Despite spatial clustering and propagation of damage, conventional research has left spatial dependencies underexplored. This study constructs graph representations that reflect component adjacency and employs a Graph Transformer to capture both local and distant dependencies. Synthetic datasets confirm the advantage of introducing spatial positioning in settings with probabilistic transitions and various component topologies. The model is also tested on a semi-automatically generated Tokyo girder bridge dataset. It improves precision sixfold over the percentage prediction method, surpasses a graph neural network, and outperforms a Transformer model without spatial information by five points on the real dataset and eight on a synthetic dataset. Attention weight analysis reveals that the model captures spatial dependencies and aligns with empirical deterioration mechanisms, offering interpretability. The proposed approach enables detailed element-level deterioration predictions, enhancing maintenance planning and safety.