Autonomous agentic AI with policy adaptation for physics-informed spectral learning in Structural Health Monitoring

This study develops a physics-grounded agentic artificial intelligence (AI) framework for Structural Health Monitoring (SHM) that integrates perception, cognition, action, and reflection with structural dynamics and stochastic system identification. The framework differs from conventional machine learning or deep learning pipelines by optimizing long-horizon monitoring objectives, autonomously adapting sensing and inference to meet safety-critical decision requirements. The methodology establishes a direct bridge from multi-degree-of-freedom structural dynamics to output-only spectral identification through Frequency Domain Decomposition (FDD), and subsequently to Bayesian inference of damage hypotheses and adaptive policy optimization. The proposed framework introduces three main contributions. First, it represents the explicit integration of agentic AI with SHM, embedding reasoning, planning, and learning within a closed-loop structure that is physically interpretable. Second, it provides a rigorous connection between computer science formulations such as partially observable Markov decision processes and reinforcement learning with the established domain of linear structural dynamics and modal analysis. Third, it demonstrates reproducibility and effectiveness through a numerical study on a five-degree-of-freedom shear building subjected to Gaussian white noise excitation. The results indicate high-fidelity modal identification and accurate localization of stiffness loss, with reliability assessed against theoretical baseline modes using the Modal Assurance Criterion (MAC). Further, the robustness of the agentic AI is validated using the ASCE SHM benchmark structure. The findings point toward a new generation of monitoring systems that are physics-informed, uncertainty-aware, and capable of resilient and adaptive operation across interconnected infrastructure networks.