FedAvg-Based CTMC Hazard Model for Federated Bridge Deterioration Assessment

A new research paper by Takato Yasuno introduces a federated learning framework that enables municipalities to collaboratively analyze bridge deterioration while maintaining data privacy. The system uses Federated Averaging (FedAvg) to train a Continuous-Time Markov Chain (CTMC) hazard model across distributed bridge inspection datasets, allowing cities to pool insights without sharing sensitive infrastructure records. This addresses a critical challenge in public infrastructure management where data governance constraints typically prevent cross-organizational data sharing, despite the clear benefits of aggregated analysis for predicting structural failures and planning maintenance.

The technical approach involves each local entity training a log-linear hazard model that tracks three key deterioration transitions (Good→Minor, Good→Severe, Minor→Severe) using covariates like bridge age, coastline distance, and deck area. During each communication round, participants upload only a compact 12-dimensional pseudo-gradient vector to a central server, which aggregates updates using momentum and gradient clipping. The framework creates a participation incentive: municipalities that contribute their local data receive periodically updated global benchmark parameters in return—insights they couldn't derive from their isolated datasets alone. This enables evidence-based lifecycle planning while preserving data sovereignty, with simulations showing consistent convergence even across heterogeneous user distributions.