Research map / Federated learning and privacy / Personalized and heterogeneous FL
Data Disparity and Temporal Unavailability Aware Asynchronous Federated Learning for Predictive Maintenance on Transportation Fleets
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DAAFL, an asynchronous federated learning method with a data-disparity and update-frequency aware aggregation weight plus a federated early stopping criterion, evaluated on oil dilution and turbofan RUL prediction.
Problem
Asynchronous federated learning for predictive maintenance on vehicle fleets suffers from data-volume disparity across clients, temporal client dropouts, and the lack of an early stopping criterion without a global validation set.
Approach
An aggregation weight per client is computed so that each client's cumulative contribution tracks its ideal share (relative data volume times update count), and the federated validation loss is aggregated with the same weights to drive asynchronous early stopping.
Key results
- On oil dilution, DAAFL achieves 1.51% MAPE versus 1.59% for LR-AFL and 1.62% for FedAvg, and the best federated U-F1 of 80.87%.
- On jet engine RUL, DAAFL achieves 9.577% MAPE and 0.088 RMSE, the best among federated models, with classic ML at 0.090 RMSE.
- DAAFL with GRU, LSTM, TCN and feed-forward models yields competitive oil dilution results across architectures, with GRU best at 1.51% MAPE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DAAFL instead of LR-AFL | Federated oil dilution prediction · Oil Dilution test set | MAPE: 1.59±0.04 -> 1.51±0.07 | Helped |
| Use DAAFL instead of LR-AFL | Federated oil dilution prediction · Oil Dilution test set | U-F1: 77.46±1.16 -> 80.87±0.59 | Helped |
| Use DAAFL instead of FedSA | Federated oil dilution prediction · Oil Dilution test set | L-F1: 89.03±00.69 -> 89.18±00.18 | No clear effect |
| Use DAAFL instead of FedSA | Federated jet engine RUL prediction · Jet Engine test set | MAPE: 10.397±3.858 -> 09.577±0.580 | Helped |
| Use DAAFL instead of Classic ML | Jet engine RUL prediction · Jet Engine test set | RSME: 0.090±0.004 -> 0.088±0.003 | Helped |
| Replace GRU with LSTM (ablation) | DAAFL oil dilution prediction · Oil Dilution test set | MAPE: 1.51±0.07 -> 1.62±0.07 | Hurt |
| Replace GRU with TCN (ablation) | DAAFL oil dilution prediction · Oil Dilution test set | MAPE: 1.51±0.07 -> 1.95±0.24 | Hurt |
| Replace GRU with Feed Forward (ablation) | DAAFL oil dilution prediction · Oil Dilution test set | MAPE: 1.51±0.07 -> 1.99±0.14 | Hurt |
| Replace GRU with LSTM (ablation) | DAAFL oil dilution prediction · Oil Dilution test set | L-F1: 89.18±0.18 -> 87.98±0.18 | Hurt |
| Replace GRU with TCN (ablation) | DAAFL oil dilution prediction · Oil Dilution test set | L-F1: 89.18±0.18 -> 83.86±1.41 | Hurt |
| Replace GRU with Feed Forward (ablation) | DAAFL oil dilution prediction · Oil Dilution test set | L-F1: 89.18±0.18 -> 87.57±0.05 | Hurt |
| Replace GRU with LSTM (ablation) | DAAFL oil dilution prediction · Oil Dilution test set | U-F1: 80.87±0.59 -> 77.65±0.59 | Hurt |
Surprising findings
- Pr-SAFL and St-AFL perform poorly; St-AFL reaches 3.34% MAPE with high variance on oil dilution.
- Baselines with higher precision achieve it by neglecting many threshold cases, leading to poor recall.
Limitations
- Dropouts are simulated with per-client patterns because no real dataset with temporal unavailability was available.
- The oil dilution dataset is private, limiting reproducibility.
- Robustness against attacks and privacy leakage from parameter updates is not addressed.
Experimental setup
Datasets: Oil dilution dataset (Volkswagen Group logistics vehicles, 23 cars) NASA turbofan engine run-to-failure simulation dataset (Saxena et al. 2008)
Benchmarks: Oil dilution dataset (private, Volkswagen) NASA turbofan jet engine RUL dataset
Metrics: MAPE RMSE L-F1 L-P L-R U-F1 U-P U-R
Baselines: Classic ML (centralised) FedAvg FedSA Pr-SAFL St-AFL LR-AFL
Builds on: FedAvg asynchronous federated learning aggregation (Xu et al. 2023) early stopping
Closest papers by meaning
- Efficient Federated Learning against Heterogeneous and Non-stationary Client Unavailability (NeurIPS 2024 · similarity 0.85)
- FADAS: Towards Federated Adaptive Asynchronous Optimization (ICML 2024 · similarity 0.84)
- FLUX: Efficient Descriptor-Driven Clustered Federated Learning under Arbitrary Distribution Shifts (NeurIPS 2025 · similarity 0.84)
- FedLMT: Tackling System Heterogeneity of Federated Learning via Low-Rank Model Training with Theoretical Guarantees (ICML 2024 · similarity 0.84)
- FLHetBench: Benchmarking Device and State Heterogeneity in Federated Learning (CVPR 2024 · similarity 0.84)
- FedCompass: Efficient Cross-Silo Federated Learning on Heterogeneous Client Devices Using a Computing Power-Aware Scheduler (ICLR 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.