剩余使用寿命(RUL)预测对于确保工业设备的安全运行和减少定期预防性维护的成本具有重大意义。然而, 对于典型的边缘设备, 其计算能力和数据存储能力有限, 较难实现设备的RUL预测, 且云和边缘之间的数据传输速率有限, 传输所有训练数据会带来较高的延迟。此外, 由于可能的利益冲突, 通常情况下很难实现所有边缘设备之间的数据共享。为此, 该文提出了一种基于联邦学习的云边协同框架。多个边缘设备和云服务器被用来训练一个基于变分自编码器(VAE)的特征提取模块和一个RUL预测模块, 无需数据共享。在每个训练周期中, 首先在所有边缘设备上使用各自的本地训练数据集训练VAE, 再将所有本地VAE上传到云端, 并根据本地训练数据的规模为所有边缘分配权重, 聚合成一个全局特征提取模块, 再发送回所有边缘设备, 以从它们的数据集中提取隐藏特征, 并将这些特征上传到云端以训练全局RUL预测器。实验结果表明：该方法可以在资源受限的条件下执行边缘设备RUL预测, 减少了数据传输延迟并能够保护数据隐私。

Objective: Predicting the remaining useful life (RUL) of industrial equipment is critical for maintaining safe operations and minimizing maintenance costs. However, RUL prediction for edge devices faces several challenges. First, edge devices often lack the computational power and storage capacity required for complex RUL prediction algorithms, making such predictions difficult. Many RUL prediction algorithms require substantial resources, which are scarce on edge devices. Second, the limited data transmission rate between the cloud and edge devices causes high latency when transmitting large data sets to the cloud, affecting real-time predictions and increasing network bandwidth usage. Additionally, data sharing among all edge devices is often impractical owing to privacy, security issues, and potential conflicts of interest, limiting models to local data and reducing their accuracy. Methods: To address these challenges, this paper proposes a cloud-edge collaboration framework for RUL prediction based on federated learning. The framework comprises two main processes. In the first process, each training device trains a variational autoencoder (VAE) using its local data set. The trained encoders are then uploaded to the cloud and aggregated using a weighted average method (FedAVG), with the number of training samples as weights. The aggregated global encoder is then downloaded to all edge devices. In the second process, the aggregated encoder extracts hidden features from the local data sets on each edge device. These features are uploaded sequentially to the cloud to train the RUL predictor. Once trained, the predictor is sent back to the edge devices, completing one training cycle. This iterative process continues until a well-trained RUL prediction model, consisting of the global encoder and predictor, is achieved. During the testing stage, the global encoder is used to extract hidden features, while the RUL predictor performs deeper feature extraction and RUL prediction. In this framework, only local encoders and hidden features are uploaded to the server, significantly reducing communication overhead. Most of the training occurs on the server, with clients only performing the basic training of the shallow VAE, thereby effectively utilizing the server's powerful computational capabilities. Data privacy is maintained since the server receives hidden features and encoders, not the original data, preventing data reconstruction. Results: To validate the proposed method's efficiency and practicality, different network structures were tested for RUL prediction on the commercial modular aero-propulsion system simulation (C-MAPSS). Although there was a slight decline in prediction performance compared to the baseline, the difference was within acceptable limits. This minor trade-off in accuracy enabled RUL prediction under resource constraints. The proposed algorithm significantly reduced data transmission time after feature extraction across various data scales consistently. In industrial scenarios with large data volumes, this reduction was even more pronounced. Further validation using nuclear power unit fault data sets showed a slight decrease in root mean square error (RMSE) on the test set without a significant drop in prediction accuracy. These results demonstrate that the proposed cloud-edge collaboration framework is promising for fault diagnosis in nuclear power units, effectively addressing edge resource limitations. Conclusions: The proposed cloud-edge collaboration framework leverages federated learning to achieve RUL prediction on resource-constrained edge devices, thereby alleviating issues related to resource constraints and data privacy. By employing VAE-based feature extraction and federated learning for model training, the framework achieves efficient model training while significantly reducing communication overhead with minimal impact on accuracy. Experimental validation on industrial simulation data sets and nuclear power unit fault data sets demonstrates the framework's practicality and effectiveness. This framework represents a useful approach to addressing challenges in fault diagnosis and URL prediction within resource-constrained settings.