Gossip training shows promise for industrial failure detection

VERA Brief

AI-generated. Grounded in the article and its cited sources.

A new study compares serverless gossip training with federated learning for failure detectors on NASA data. Gossip training shows promise as a viable decentralized approach for industrial predictive maintenance, transmitting similar data payloads to federated averaging without a central coordinator.

Key facts

• Serverless gossip training was evaluated as an alternative to federated learning for LSTM-based failure detectors.
• Gossip training achieved an F1 score of 89.6% +/- 1.3% on the FD001 dataset, closely matching federated averaging.
• Peer communication in both gossip and federated learning improved F1 scores compared to local training across FD002-FD004.
• Gossip learning matched federated averaging performance on FD003 and FD004.
• Simulated message loss and node or server outages did not significantly impact the results of gossip training.