马新娜, 李豪.基于动态聚类联邦学习的轴承故障诊断方法[J].河南理工大学学报（自然科学版）,2025,44(5):1-8.

MA X N, LI H. Bearing fault diagnosis method based on dynamic clustering federated learning [J]. Journal of Henan Polytechnic University (Natural Science), 2025, 44(5): 1-8.

基于动态聚类联邦学习的轴承故障诊断方法

马新娜, 李豪

石家庄铁道大学 信息科学与技术学院，河北 石家庄 050043

摘要: 目的 工业生产中，各个客户端持有的轴承故障数据大多服从非独立同分布（non-independent and identically distributed，Non-IID），为了解决联邦系统因非独立同分布数据导致的故障诊断性能不佳问题开展相关研究。  方法 提出基于动态聚类联邦学习的轴承故障诊断方法（dynamic clustering federated learning，DCFed），该方法将客户端分组到具有相似数据分布的集群中，缓解数据异构导致的模型偏移问题，它包含参数选择阶段、聚类阶段和聚合阶段。首先为客户端提供最佳的训练模型参数，然后通过计算本轮训练中客户端之间的推理相似性（余弦相似度）进行K-means聚类，从而使每个聚类簇中的客户端尽量服从于独立同分布模式，最后将同一集群中的客户端进行平均聚合，再将不同集群聚合，得到高效和个性化的联邦学习模型。  结果 在凯斯西储大学（CWRU）轴承数据集上进行不同维度测试，所提方法在Non-IID场景下的全局模型准确率为95.23%，并且接近于集中训练（Central）方法，且只需150个训练轮次就达到收敛，与其他方法相比有着更快的收敛速度和更高的准确率，最后探究关键参数及聚类方法的影响，结果表明，所提方法有着更强的鲁棒性和稳定性。  结论 在非独立同分布条件下，该方法能够有效地减少数据异构对于联邦学习的影响，加快收敛速率。不同实验结果表明该方法能够出色完成故障诊断任务。

关键词:数据异构;聚类联邦学习;故障诊断;深度学习;数据隐私

DOI: 10.16186/j.cnki.1673-9787.2024080007

基金项目:国家自然科学基金资助项目（12172234）；河北省自然科学基金资助项目（A2021210022）；河北省三三三人才科研项目（A202101018）

收稿日期:2024/08/08

修回日期:2024/09/23

出版日期:2025/07/23

Bearing fault diagnosis method based on dynamic clustering federated learning

Ma Xinna, Li Hao

School of Information Science and Technology， Shijiazhuang Tiedao University， Shijiazhuang 050043， Hebei， China

Abstract: Objectives In industrial scenarios， bearing fault data held by clients are typically non-independent and identically distributed （Non-IID）， which adversely affects the performance of federated learning （FL） systems. This study aimed to address the performance degradation of FL under such conditions. Methods This paper proposed a bearing fault diagnosis method based on dynamic clustering federated learning （DCFed）. The method grouped clients with similar data distributions into clusters to mitigate model divergence caused by data heterogeneity. It comprised three main stages： parameter selection， clustering， and aggregation. In the parameter selection stage， optimal model parameters were provided to the clients. During clustering， K-means clustering was performed based on cosine similarity calculated from client inference results within the current training round， encouraging clients within each cluster to approximate an independent and identically distributed （IID） pattern. In the aggregation stage， local models were first averaged within clusters and then aggregated across clusters to form an efficient and personalized global model. Results The proposed DCFed method was evaluated on the Case Western Reserve University （CWRU） bearing dataset. Under the Non-IID scenario， the global model achieved an accuracy of 95.23%， which was close to that of centralized training methods. DCFed converged within just 150 training rounds， demonstrating faster convergence and higher accuracy compared to other methods such as FedProx and Fedora. Further experiments examining key parameters and clustering methods confirmed the robustness and stability of the proposed approach. Conclusions Under Non-IID conditions， the proposed method effectively mitigated the impact of data heterogeneity on federated learning and accelerates convergence. Experimental results demonstrated that the method performed excellently in bearing fault diagnosis tasks.

Key words: data heterogeneity; clustering federated learning; fault diagnosis; deep learning; data privacy