LLM-Powered Report-Driven Markov Modelling for Large-Scale Predictive Bridge Maintenance in Japan

Reliable bridge management depends on timely, data-driven insight into how individual components deteriorate, yet much of the evidence resides in narrative inspection cards, photographs, and semi-structured PDFs that are still keyed by hand. We present a large-language-model (LLM) pipeline that automatically parses these documents, extracting per-component condition ratings, damage mechanisms (e.g., cracks, corrosion), repair actions, and traffic statistics into an analysis-ready warehouse. From the harvested time series, we estimate a non-stationary, mechanism-aware Markov model whose transition intensities are conditioned on heavy-vehicle flow, capturing both monotone ageing and post-repair recovery. Closed-form propagation yields twenty-year condition distributions, exceedance risks, and life-cycle-cost (LCC)–optimal intervention years without Monte Carlo simulation or reinforcement learning. On a corpus of 800 Japanese bridges (10,000) components), the extraction stage achieves near-human accuracy and eliminates manual coding, while the resulting component-resolved forecasts reduce expected LCC by 18% and severe-state risk by 23% compared with periodic schedules and rank-only baselines. The approach scales linearly with the number of reports and preserves full interpretability, providing asset owners with transparent, auditable metrics that can be used directly for predictive maintenance and budget planning.