Title: Robust Infrastructure Inspection via Hierarchical Federated Learning Incorporating Dynamic Clustering and Adaptive Regularization

Abstract: The integration of data-driven computer vision models into structural health monitoring (SHM) is significantly impeded by data silos, a consequence of strict privacy and security mandates. Although federated learning (FL) presents a privacy-centric collaborative solution, its scalability across national infrastructure networks is limited by the problem of "double heterogeneity." This issue arises from macro-level physical differences between varied structural types and micro-level statistical disparities within local datasets. To address these obstacles, we introduce a novel hierarchical federated learning framework featuring a synergistic two-tier optimization approach. At the macro level, a dynamic gradient-based clustering system automatically groups distributed clients into specialized expert categories according to their structural degradation patterns, eliminating the dependency on pre-existing geographical data. Simultaneously, at the micro level, an intra-cluster Dynamic Region-Adaptive Proximal Regularization (DRAPR) component calculates a real-time Non-IID Intensity Score for each participant. By adjusting a proximal penalty in response to local label skewness and gradient divergence, DRAPR optimizes local model updates, reduces client drift, and safeguards against the catastrophic forgetting of less frequent damage types. Extensive testing on a large-scale, real-world structural inspection dataset confirms that combining macro-clustering with micro-regularization effectively resolves dual-level heterogeneity, resulting in highly robust and specialized diagnostic models suitable for complex infrastructure assessment.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC