Clustered Federated Learning for Population-Based Structural Health Monitoring

This article deals with two major challenges in population-based structural health monitoring (PBSHM) for critical infrastructures, with focus on bridges. The first challenge is related to training a collaborative model based on local datasets from different bridges under privacy-preserving constraints. The second challenge is related to the (potential) limited amount of data and/or resources available locally at some bridges. To address these issues, we propose a framework based on clustered federated learning (CFL) for efficient and collaborative training of monitoring models while preserving data privacy. The approach leverages the concept of principal angles (PAs) to cluster the bridges according to their data distributions without domain-based prior information. Cluster-specific models are then trained efficiently according to the available local resources. Moreover, the proposed framework allows the integration of new bridges providing quick and resource-efficient knowledge transfer from the cluster-specific parameters to the infrastructure-specific parameters. Numerical results based on extensive simulations show that the proposed framework performs well, both within supervised and unsupervised settings, yielding more accurate results than traditional schemes.