Graph Neural Networks combined with PCA for predicting blast load time series on structures

A comprehensive understanding of blast load information on structures is crucial for the assessment of their safety and reliability. While the predictive efficacy of machine learning in blast load is well-established, traditional Artificial Neural Networks (ANN) encounter challenges in extracting intricate features that characterize the interaction between structures with complex geometry and blast loads, as well as their spatial distribution. A novel method, PCA-GNN, is proposed for predicting blast load time series on structures by integrating Graph Neural Network (GNN) and Principal Component Analysis (PCA). The main idea of this model involves preprocessing the time series data using PCA to extract essential features, while employing GNN to capture the spatial correlation of blast loads on structures. The blast load data obtained from the box girder with a complex cross-section is utilized to evaluate the performance of PCA-GNN and PCA-ANN models. The results indicate that the PCA-GNN model substantially outperforms the PCA-ANN model in accurately replicating blast load characteristics, encompassing overpressure time series and the peaks of overpressure and impulse. Consequently, the PCA-GNN model demonstrates promising potential for predicting blast loads on structures, contributing to a more efficient and effective risk management strategy that enhances the reliability and safety of explosion-related hazards.