Federated Estimation of Infrastructure Deterioration

Assessing the health of public infrastructure, such as bridges, relies on periodic inspections containing sensitive data. Sharing such data between organizations is often impractical due to governance and privacy constraints.

This article presents a federated approach for estimating bridge deterioration, based on Continuous-Time Markov Chains (CTMC). The framework allows different municipalities to jointly train a shared benchmark model without the need to transfer raw inspection data.

Model Operation

Each local entity retains its inspection data and trains a CTMC log-linear hazard model, considering deterioration transitions such as: Good → Minor, Good → Severe, and Minor → Severe. Covariates include bridge age, coastline distance, and deck area.

Local optimization occurs via mini-batch stochastic gradient descent on the CTMC log-likelihood. Only a 12-dimensional pseudo-gradient vector is uploaded to a central server for each communication round. The server aggregates updates using sample-weighted Federated Averaging (FedAvg) with momentum and gradient clipping.

Participation Incentives and Data Sovereignty

The federated update mechanism provides a natural incentive to participate: entities that register their inspection datasets on a shared technical-standard platform receive periodically updated global benchmark parameters in return. This information would not be obtainable from local data alone, enabling evidence-based life-cycle planning without compromising data sovereignty.

For those evaluating on-premise deployments, there are trade-offs to consider. AI-RADAR offers analytical frameworks on /llm-onpremise to evaluate these aspects.