Efficient Training of Large-Scale Industrial Fault Diagnostic Models through Federated Opportunistic Block Dropout

• DOI: 10.1609/aaai.v37i13.26836
• Publication Date: 2023-06-27T18:32:17Z
• Authors: Yuanyuan Chen; Zichen Chen; Sheng Guo; Yansong Zhao; Zelei Liu; Pengcheng Wu; Chengyi Yang; Zengxiang Li; Han Yu
• Keywords: Leverage (statistics); Upload; Acknowledgement; Dropout (neural networks)

Artificial intelligence (AI)-empowered industrial fault diagnostics is important in ensuring the safe operation of applications. Since complex systems often involve multiple plants (possibly belonging to different companies or subsidiaries) with sensitive data collected and stored a distributed manner, collaborative diagnostic model training needs leverage federated learning (FL). As scale models are large communication channels such not exclusively used for FL training, existing deployed frameworks cannot train efficiently across institutions. In this paper, we report our experience developing deploying Federated Opportunistic Block Dropout (FedOBD) approach training. By decomposing large-scale into semantic blocks enabling participants opportunistically upload selected quantized it significantly reduces overhead while maintaining performance. its deployment ENN Group February 2022, FedOBD has served two coal chemical cities China build prediction models. It helped company reduce by over 70% compared previous AI Engine, performance at 85% test F1 score. To knowledge, first successfully dropout-based approach.