Collaborative System Failure Prognostics via Federated Longitudinal-Survival Modeling

TL;DR

The paper introduces a federated learning framework that enables collaborative failure prognosis across distributed systems without sharing raw sensor data or failure times. It uses a separable discrete-time hazard model combined with longitudinal representation learning to estimate system reliability and remaining useful life. The approach achieves performance close to centralized training while respecting data privacy across diverse operational conditions.

Why It Matters

Many industrial systems generate sensitive operational data that cannot be shared centrally, limiting the development of accurate predictive maintenance models. This work enables organizations to jointly improve prognostic accuracy without compromising data privacy, which is critical for real-world deployment in aerospace, energy, and manufacturing sectors.

Key Findings

• The proposed model uses a client-separable hazard objective, allowing federated training without requiring global risk set computation.
• Longitudinal representations from multivariate sensor data effectively capture system degradation patterns across different failure modes.
• Performance across four C-MAPSS engine datasets matches centralized training results under decentralized settings, even with heterogeneous operating conditions.