Label synchronization strategies for hybrid federated learning

Advances in artificial intelligence and sensing technologies are becoming instrumental in the adoption of condition-based and predictive maintenance strategies across various industrial sectors. Data-driven models enable the estimation of the current health state of systems and forecast their temporal evolution, promoting proactive maintenance approaches. Despite the evident benefits of collaborative data sharing for developing these models, challenges persist, including data scarcity in each stakeholder, the diversified nature of information held by various parties, and privacy concerns that hinder the willingness to disclose sensitive data. Consequently, parties monitoring similar systems have potential interest in collaborative solutions that respect data privacy policies. Federated Learning, a privacy-preserving and decentralized machine learning approach can address these challenges. By introducing various data partitioning settings – horizontal, vertical, and hybrid – it allows for different modes of data contribution from involved parties, thereby determining how data synchronization occurs. In this work, we introduce FedMulLabSync, a novel algorithm that extends the label synchronization to multidimensional spaces to address use cases where sharing unidimensional labels is insufficient. FedMulLabSync incorporates two synchronization strategies: time-sharing and similarity-based. Time-sharing augments the label space with time, while similarity-based computes the similarity between two multidimensional target labels of different parties. Both strategies are inspired by solutions designed to overcome non-independent and identically distributed (non-IID) data. The effectiveness of the proposed algorithm is demonstrated in forecasting and classifying the health condition of two systems using labels in multidimensional space: turbofan and maritime engines. In forecasting the remaining useful lifetime of turbofan engines, the MAE of the most precise non-federated model was reduced from ≈ 12 . 58 to ≈ 9 . 97 . In classifying the health condition of maritime engines, the F-score of the worst non-federated model was increased from 0.22 to 0.45. • Label synchronization in hybrid federated learning applications. • Strategies for handling non-IDD data heterogeneity in hybrid federated learning. • Client-server architecture for exchanging diverse information in federated solutions. • Challenges in collaborative prognostics and collaborative prognosis. • Collaborative RUL prediction of turbofans and anomaly detection in ship engines.