From Multi-Source Monitoring to Trustworthy Decisions: Hybrid AI-Physics StateSpace Models for Transportation Infrastructure

Transportation infrastructure monitoring is increasingly characterized by heterogeneous, multi-rate data streams spanning in-situ IoT sensing (e.g., accelerations, strains, inclinations, temperatures), geodetic measurements, and remote sensing (e.g., InSAR displacement time series). Converting such observations into uncertainty-aware, decision-support-oriented information remains challenging due to missingness, asynchronous sampling, nonstationary environmental effects, and the need to propagate uncertainty from measurement to maintenance action. This paper proposes a hybrid AI-physics state-space framework that unifies representation learning, probabilistic data assimilation, and unsupervised regime discovery for network-scale asset management.
A masked, self-supervised encoder first converts heterogeneous sensor and remote-sensing observations into a compact latent representation that can handle missing data and imperfect alignment between sources. These latent measurements are then fused using an asynchronous, multi-rate Kalman Filter (KF) defined on a physically interpretable latent state representation, where dynamic features such as displacement trends and velocity-like behaviour are implicitly encoded by the neural encoder, so that InSAR and in-situ sensors can be combined in a dynamically consistent way. To account for changing behavior over time and hidden deterioration patterns, the KF state trajectories are grouped using Bayesian non-parametric mixture models, which can discover the appropriate number of regimes automatically and assign each time period a probability of belonging to each regime. Overall, the framework outputs uncertainty-aware state estimates, forecasts, and regime probabilities, enabling probabilistic alarms and decision rules (e.g., exceedance probabilities and confidence-weighted warnings) instead of fixed threshold triggers. Collectively, the framework advances the state of practice from "monitoring and visualization" toward uncertainty-calibrated, explainable decision support for transportation infrastructure.

Keywords: Hybrid AI-physics modeling, state-space models, InSAR displacement time series, Kalman Filter (KF), Bayesian non-parametric clustering, uncertainty quantification.