Memory-augmented spatio-temporal graph networks for federated cross-domain remaining useful life prediction

Federated learning (FL) enables multiple industrial clients to jointly train remaining useful life (RUL) prediction models without sharing raw data, which fits industrial scenarios with local data confinement requirements. Sensor data collected from different clients under varied operating conditions exhibit heterogeneous distributions, which hinder cross-domain knowledge transfer, weaken degradation feature alignment, and further degrade federated RUL prediction performance. To tackle this challenge, this paper proposes a prognostic architecture termed Memory-Augmented Spatio-Temporal Hierarchical Graph Network with Domain Adaptation (MA-STHGN-DA) for privacy-preserving federated prognostics. Specifically, a Memory-Augmented Spatio-Temporal Hierarchical Graph Network (MA-STHGN) is developed to model dynamic dependencies of multivariate sensor signals and capture long-term degradation via hierarchical graph learning and memory-enhanced feature extraction. On this basis, a decoupled dual-memory domain adaptation mechanism is introduced to align spatio-temporal graph representations between the labeled source and unlabeled target domains by segregating domain-invariant degradation physics from condition-specific local dynamics, which mitigates distribution shifts while preventing negative transfer. Furthermore, the standard federated averaging algorithm is integrated with a reputation module to construct Federated Averaging with Reputation (FedAvg-Rep) for decentralized training. This deployment-oriented strategy assigns aggregation weights according to local prediction performance and alignment quality, thereby reducing the influence of unreliable client updates. Experiments and ablation studies show that the proposed method consistently improves the accuracy of cross-domain RUL prediction while preserving local data privacy, demonstrating its effectiveness in distributed prognostics.