主动学习解耦注意残差网络的轴承复合故障诊断  

Composite Fault Diagnosis of Bearing Based on Active Learning Decoupling Attention Residual Network

作　　者：李春亚[1] 陈晨[2] LI Chun-ya;CHEN Chen(Zhengzhou Railway Vocational&Technical College,He'nan Zhengzhou 451460,China;He'nan University of Technology,He'nan Zhengzhou 450001,China)

机构地区：[1]郑州铁路职业技术学院,河南郑州451460 [2]河南工业大学,河南商郑州450001

出　　处：《机械设计与制造》2024年第11期189-197,共9页Machinery Design & Manufacture

基　　金：河南省高等学校重点科研项目(17A460031)。

摘　　要：考虑到现有的深度学习方法通常需要大量的标记数据,在实际应用中难以实现,提出了一种基于主动学习解耦注意残差网络的轴承复合故障诊断方法。首先利用主动学习技术从大量未标记数据中选择一些最有利的数据来提高模型性能,降低了对标记复合故障数据的要求。然后另外将注意模块与残差块相结合,提出了一种基于多标签熵的特征选择策略,以获取模型中最有用的未标记数据,并对这些数据进行标记。进一步将多标签解耦分类器代替常用的softmax分类器,使模型具有更好的复合故障识别能力。在轴承数据集上的实验结果证明提出方法在保证最终模型性能的前提下,能够大大减少复合故障标注的工作量。Considering that the existing deep learning methods usually need a large amount of labeled data and are difficult to realize in practical application,a bearing composite fault diagnosis method based on active learning decoupling attention residual network was proposed.Firstly,active learning technology was used to select some of the most favorable data from a large number of unlabeled data to improve the model performance and reduce the requirements for labeled composite fault data.Then,combining the attention module with the residual block,a feature selection strategy based on multi label entropy was proposed to obtain the most useful unlabeled data in the model and label these data.Furthermore,the multi label decoupling classifier was used to replace the commonly used softmax classifier,so that the model had better compound fault recognition ability.The experimental results on the bearing data set show that the proposed method can greatly reduce the workload of compound fault annotation on the premise of ensuring the performance of the final model.

关 键 词：主动学习 解耦注意残差网络 轴承 复合故障诊断 

分 类 号：TH16[机械工程—机械制造及自动化] TH133.33[自动化与计算机技术—控制理论与控制工程] TP18[自动化与计算机技术—控制科学与工程]