Serverless Gossip Training vs. Federated Learning for LSTM-based Failure Detection: A Comparative Study
Summary:This research compares the performance of serverless gossip training with federated averaging (FedAvg) for LSTM-based failure detection models using the NASA C-MAPSS turbofan dataset. Results show that gossip training achieves comparable performance to FedAvg (89.6% vs. 89.9% F1 on FD001) without requiring a central server, outperforming isolated local training (83.6%). The study also demonstrates the robustness of gossip training against simulated message loss, node failure, and server outages.
Background and Motivation
Industrial predictive maintenance increasingly relies on learning from equipment across different sites, where pooling sensor data is challenging. While federated averaging (FedAvg) addresses this with a central server, gossip learning offers a serverless alternative. This study evaluates the performance of gossip training for LSTM-based failure detection models and compares it with FedAvg, isolated local training, and centralized training.
Methodology and Experiments
The experiments were conducted on the NASA C-MAPSS turbofan dataset. All methods shared the same implementation, architecture, initialization, optimizer, data split, and training budget. The primary evaluation metric was the terminal window F1 score to avoid statistical dependence from overlapping windows.
Key Findings
- Performance Comparison: On the FD001 subset, gossip training achieved an F1 score of 89.6%, comparable to FedAvg's 89.9%, and significantly outperformed local training (83.6%) and centralized training (93.5%).
- Communication Efficiency: Gossip training transmitted the same payload as FedAvg without a coordinator.
- Node Agreement: Node models agreed closely but not exactly (1.8% pairwise decision disagreement vs. 5.6% without communication).
- Multi-Condition Subset Performance: On FD002-FD004 subsets, peer communication improved terminal window F1 by 13-28 points; gossip matched FedAvg on FD003 and FD004 but was 4.3 points lower on the multi-condition FD002 subset.
- Robustness: Simulated message loss, node failure, and server outages did not significantly affect either method, but larger gossip rings degraded performance faster.
Conclusion and Recommendations
Gossip training is a practical serverless alternative when data heterogeneity is moderate. As heterogeneity increases, faster-mixing topologies become crucial. The study demonstrates that gossip training not only matches FedAvg in performance but also offers better scalability and robustness, providing new insights for distributed learning in industrial predictive maintenance.
Developer Recommendations
- Use Cases: Suitable for scenarios where data heterogeneity is moderate and centralized storage is not feasible.
- Optimization Directions: Explore faster-mixing topologies to enhance gossip training performance in highly heterogeneous data environments.
- Security: Design gossip training systems with considerations for node failures and communication interruptions.
— END —Source: ArXiv Machine Learning (cs.LG) (2026-09-30)
Tags: #LSTM #Gossip Training #Federated Learning #Failure Detection #Distributed Learning
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments