Speaker
Description
Conventional bridge deterioration prediction models do not account for the spatial relationships between structural components and damage locations. Since deterioration is a spatially distributed phenomenon, incorporating positional relationships can lead to more accurate predictions.
In previous research, the authors developed a Markov chain-based deterioration prediction model incorporating Graph Transformer architecture to capture detailed spatial relationships between bridge components. Based on the research, this study focuses on positional embedding, a crucial factor influencing Transformer model performance, and evaluates its impact through comparative experiments.
Specifically, two approaches were examined at first: one utilizing adjacency degree in a graph representation, and another directly embedding the coordinates of representative points in structural drawings. Both approaches outperformed a percentage prediction method and a Transformer without positional encoding. However, in terms of a precision metric about the presence of damage progression, the coordinate-based method achieved a significantly higher score (81.7%) compared to the graph-based method (43.3%).
Additionally, for the coordinate-based approach, a comparative analysis of different positional embedding architectures was conducted, including a simple multilayer perceptron (MLP) and a mechanism inspired by PointNet that considers all components in the data to capture relative spatial relationships.
This study contributes to the development of a highly reliable deterioration prediction model, enabling the optimization of inspection frequency based on degradation conditions and facilitating more efficient maintenance and management of bridge infrastructure.
