基于自适应聚类与预聚合的联邦学习故障诊断方法研究  

Research on Federated Learning Fault Diagnosis Method Based on Adaptive Clustering and Pre-aggregation

作　　者：高聪 谢怡宁 石江涛[1] Gao Cong;XieYining;Shi Jiangtao(School of Computer and Control Engineering,Northeast Forestry University,Harbin 150040,China;School of Mechanical and Electrical Engineering,Northeast Forestry University,Harbin 150040,China)

机构地区：[1]东北林业大学计算机与控制工程学院,哈尔滨150040 [2]东北林业大学机电工程学院,哈尔滨150040

出　　处：《机电工程技术》2024年第8期41-46,共6页Mechanical & Electrical Engineering Technology

基　　金：黑龙江省科技厅省级重点研发计划指导项目(GZ20220088);黑龙江省科技厅省重点研发计划应用研究项目(SC2022ZX06C0025);哈尔滨市科技局制造业创新人才项目(CXRC20221110393)。

摘　　要：联邦学习在旋转机械故障诊断领域有着很好的发展前景,然而面对数据异构的问题,现有方法准确性严重下降。对此,提出了一种基于自适应聚类与预聚合的联邦学习框架(FLACPA)。通过改进遗传算法,实现了无需预设簇类的自适应聚类,从而得到最优的数据集群划分,此外,该框架还引入了预聚合的策略,各数据集群同时进行预聚合,通过在中介服务器上训练局部模型,有效地存储了局部信息,为联邦学习的发展提供了新的思路。在两个数据集上的实验结果表明,所提方法相较于之前联邦学习框架准确率均有提升,准确率平均提升了约2.75%。同时,该方法降低了模型训练对计算资源的需求,收敛时间最高提升了43%,显著提高了模型训练的效率。这一创新性的联邦学习框架为旋转机械故障诊断领域提供了一种有效而可行的解决方案。Federated learning shows promising prospects in the field of rotating machinery fault diagnosis.However,facing the challenge of data heterogeneity,existing methods suffer from a severe decline in accuracy.To address this issue,a federated learning framework based on adaptive clustering and pre-aggregation(FLACPA)is proposed.By improving genetic algorithms,adaptive clustering without presetting clusters is achieved,optimizing data set partitioning.Additionally,the framework introduces a pre-aggregation strategy,where each data cluster is pre-aggregated simultaneously.By training local models on intermediary servers,local information is effectively stored,offering a novel perspective for the development of federated learning.The experimental results on two datasets show that the proposed method improves accuracy compared to the previous federated learning framework,with an average accuracy improvement of about 2.75%.Meanwhile,this method reduces the computational resource requirements for model training,with a maximum convergence time improvement of 43%,significantly improving the efficiency of model training.This innovative federated learning framework provides an effective and feasible solution for the field of rotating machinery fault diagnosis.

关 键 词：故障诊断 联邦学习 自适应聚类 预聚合 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]