Causal inference-based graph neural network method for predicting asphalt pavement performance

To enhance the prediction accuracy of asphalt pavement rutting, this study introduces an end-to-end multivariate time series prediction model that integrates graph neural networks(GNN) with causal inference methodologies.The proposed model aims to effectively capture long-term and short-term temporal patterns as well as interdependencies among multiple variables. The model comprises four modules: global feature extraction, local feature extraction,causal inference, and dual-channel graph convolution. The global feature extraction module employs attention mechanisms and gated recurrent units(GRU) to capture long-term temporal dependencies within variables. The local feature extraction module utilizes dilated convolutional neural networks(CNN) with various kernel sizes to extract short-term temporal patterns at different scales. In the causal inference module, relationships among variables are identified using transfer entropy based on information theory, resulting in a relationship coefficient matrix that quantifies complex dependencies among variables. The dual-channel graph convolution module extends traditional low-pass graph convolutional neural networks by integrating a high-pass filter, simultaneously capturing low-frequency and high-frequency components of node signals or features to potentially improve prediction accuracy. The proposed approach was evaluated using the RIOHTrack dataset from the Research Institute of Highway Track, with comparisons conducted against several benchmark models, including the classical statistical model VARIMA, shallow learning model SVR, deep learning model GRU, attention mechanism-enhanced GRU, and TE-GCN. Experimental results indicate that the proposed model achieves superior predictive performance across various categories of asphalt pavement structures. Compared to traditional statistical models, deep learning-based models are more effective and stable, and the GRU module enhanced with attention mechanisms can capture long-term dependencies, further enhancing predictive performance. Overall, the proposed model provides a potentially effective solution for predicting asphalt pavement rutting and may offer practical insights for future pavement structure design and maintenance planning aimed at extending pavement lifespan.