Multisource Data Fusion Approach for Predicting the Deterioration of Sign Structures along Highways

Sign structures along highways are essential components of the road transportation infrastructure system by providing useful information to drivers, but they are susceptible to fatigue cracking and corrosion. Accurately forecasting their deterioration enables transportation agencies to prioritize inspections and optimize maintenance budgets. In this paper, we develop a multisource data fusion approach based on autoencoder for predicting the deterioration of sign structures. The proposed model integrated multiple autoencoders to extract meaningful representations from raw data while the latent representation from each autoencoder corresponded to one head of the neural network and was independently processed on a multilayer perceptron (MLP). Furthermore, self-attention was incorporated into each head of the network to assign different weights, which allows the model to capture informative features and patterns within the data while downplaying the less relevant ones. In our experiment, structural attributes, traffic, and environmental data were collected from multisource databases to consider the various factors contributing to the fatigue cracking and corrosion of sign structures. Experimental results demonstrate that the proposed model achieves an AUC of 0.814 in predicting sign structure deterioration, outperforming the standard MLP (0.759), random forests (0.788), and XGBoost (0.794). Feature importance analysis using Shapley additive explanations (SHAP) identifies structure age and truck volume as the most influential predictors, with nine of the top fifteen features being environmental variables, underscoring the critical role of climatic stressors in structural degradation. By predicting sign structure deterioration, transportation agencies can implement proactive maintenance strategies and life cycle cost planning, ensuring the integrity and reliability of sign structures and enhancing road transportation safety.