CausalPD: Joint Causal Discovery and Intervention for Large-Scale Pavement Distress Distribution Data.

Abstract

Large-scale pavement distress distribution modeling is vital for optimizing pavement inspection systems (PISs), preventive maintenance, and overall infrastructure resilience. Generic urban computing, correlation-based, or spatiotemporal methods exhibit significant performance limitations in capturing the intricate degradation processes of pavements, especially under sparse, nonstationary data conditions due to infrequent inspections. Moreover, specialized approaches focusing on annual or sectional deterioration frequently lack the granularity required for daily maintenance decision-making. In response, we propose CausalPD, a novel framework that integrates joint causal discovery and intervention to predict pavement distress distributions. CausalPD employs a transformer-based architecture to extract critical causal patches from historical data, effectively leveraging the interaction between intrinsic degradation patterns and external environmental factors. By modeling causal relationships and mitigating confounding influences, our approach distinguishes genuine causal signals from noise, thereby enhancing both prediction accuracy and interpretability. Extensive experiments on real-world datasets demonstrate that CausalPD outperforms state-of-the-art methods across diverse scenarios, forecasting horizons, and spatial network configurations. Our comprehensive analysis validates the model's robust inductive biases and generalization capabilities, establishing CausalPD as a promising causal modeling paradigm for infrastructure health monitoring and proactive maintenance planning. PyTorch implementation and datasets are available at https://github.com/xuesong-wu/CausalPD.