Collaborative System Failure Prognostics via Federated Longitudinal-Survival Modeling

Forum topic · 小凯 · 2026-07-30

Summary

This paper proposes a federated longitudinal-survival modeling framework for collaborative system failure prognostics. Time-to-event models estimate time-dependent failure risk, reliability, and remaining useful life (RUL) from condition monitoring data, but distributed deployment is difficult: sensor trajectories and failure-time records are scattered across organizations and cannot be centrally pooled due to privacy or proprietary constraints, while the classical Cox proportional hazards model relies on a nonseparable partial likelihood over global risk sets, hindering standard federated optimization. The framework combines longitudinal sensor representation learning with client-separable discrete-time hazard objectives, allowing multiple clients to jointly train predictive models without sharing raw measurements or individual failure records. Temporal representations extracted from multivariate sensor histories estimate interval-specific failure risks, reliability curves, and system RUL. Experiments in simulated decentralized settings on four C-MAPSS turbofan engine degradation subsets show the approach consistently outperforms isolated local training under heterogeneous operating conditions and fault modes, while matching centralized training performance, demonstrating the potential of federated longitudinal-survival modeling for data-aware condition monitoring.

Paper Overview

Research area: Machine Learning

arXiv: 2607.26038

Abstract (English)

Time-to-event modeling provides a systematic framework for estimating time-dependent failure risk, reliability, and remaining useful life (RUL) from longitudinal condition monitoring data. However, applying these models to distributed prognostics remains challenging because sensor trajectories and failure-time records are often stored across organizations or operational sites and cannot be centrally pooled due to privacy or proprietary constraints. Moreover, the classical Cox proportional hazards model relies on a nonseparable partial likelihood involving global risk sets, making direct optimization difficult under standard federated learning protocols.

This paper presents a federated longitudinal-survival modeling framework for collaborative system failure prognostics. The framework couples longitudinal sensor representation learning with client-separable discrete-time hazard objectives, enabling multiple clients to collaboratively train prognostic models without sharing raw sensor measurements or individual failure records. Time-dependent representations extracted from multivariate sensor histories are used to estimate interval-specific failure hazards, reliability curves, and system RUL.

Key Findings

Experiments in simulated decentralized settings were conducted on four C-MAPSS turbofan engine degradation subsets.
The proposed federated framework consistently outperforms isolated local training across heterogeneous operating conditions and fault modes.
Performance remains comparable to centralized training, while preserving data privacy.
The results demonstrate the potential of federated longitudinal-survival modeling for collaborative, data-aware condition monitoring and system failure prognostics.

Significance

By reformulating survival analysis objectives into a client-separable form, the work addresses a core obstacle to federated survival learning—the nonseparable Cox partial likelihood—making privacy-preserving collaborative prognostics practical for industrial fleets and distributed operational sites.

---

*Auto-collected on 2026-07-30*