Federated Learning for Aircraft Engine Prognostics Under Client Heterogeneity
A recent arXiv preprint (2608.04045) explores the themes of robust and personalized federated learning specifically for aircraft-engine prognostics, tackling both benign and adversarial client diversity. This research, carried out by a team of scientists, aims to develop remaining-useful-life (RUL) models utilizing engine sensor telemetry while avoiding the exchange of raw data. Benign heterogeneity involves honest operators experiencing various operating conditions and fault modes, whereas adversarial heterogeneity pertains to compromised operators providing tainted updates. The study employs a multi-task one-dimensional convolutional neural network and a non-IID partition from the Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) benchmark. It compares four solutions for benign heterogeneity and assesses five attacks against four aggregation methods, including a sensor-value backdoor designed to obscure engine degradation. Findings indicate that shared-representation personalization reduces the local-to-centralized root-mean-square-error gap by roughly 70%. The research prioritizes safety-oriented evaluations and controlled experiments.
Key facts
- The paper is arXiv:2608.04045, announced as a cross-type preprint.
- Federated learning enables joint training of RUL models without sharing raw sensor data.
- Benign heterogeneity stems from different operating conditions and fault modes among honest operators.
- Adversarial heterogeneity involves compromised operators submitting poisoned updates.
- The model used is a multi-task one-dimensional convolutional neural network.
- The benchmark is C-MAPSS, partitioned in a structurally non-IID manner.
- Four remedies for benign heterogeneity and five attacks against four aggregation methods are evaluated.
- Shared-representation personalization closes about 70% of the local-to-centralized RMSE gap.
Entities
—