摘要
针对变工况条件下滚动轴承故障诊断模型泛化性能不佳的问题,基于深度域自适应与半监督学习技术,提出一种带有辅助分类器的半监督卷积神经网络(semi-supervised convolutional neural network based on auxiliary classifier,简称SSCNN-AC)滚动轴承跨域故障诊断模型。首先,为提升训练过程中目标域样本伪标签的置信度,所提模型引入最近邻中心分类器作为辅助分类器,以类中心与样本嵌入特征间的余弦距离为目标域样本生成伪标签,有效提升伪标签的可靠性;其次,采用带有标签平滑项的交叉熵损失函数计算分类损失,抑制伪标签噪声对半监督学习的不利影响,提升模型泛化性能;最后,以2个不同数据集的实验结果分析对所提模型进行验证。结果表明:所提模型可有效对齐不同工况下振动信号的嵌入特征,在滚动轴承的跨域故障诊断方面具有明显优势。
To address the limited generalization performance of rolling bearing fault diagnosis models under varying operating conditions,a cross-domain fault diagnosis model with semi-supervised convolutional neural network based on auxiliary classifiers is proposed.First,to enhance the reliability of pseudo-labels for target domain samples during training,the model introduces a nearest neighbor center classifier that generates pseudolabels according to the cosine distance between class centers and sample embedding features,which effectively improves the reliability of the pseudo-labels.Second,the classification loss is computed using a cross-entropy function that includes a label-smoothing component,which mitigates the adverse effects of pseudo-labels noise on semi-supervised learning and strengthens the generalization capability of the model.Finally,experimental analyses on 2 different datasets were conducted to validate the proposed method.The results indicate that the proposed model effectively aligns the embedding features of vibration signals under different operating conditions,and shows the significant advantages in cross-domain fault diagnosis of rolling bearings.
作者
何天经
赵荣珍
魏孔元
董晓鑫
邓林峰
HE Tianjing;ZHAO Rongzhen;WEI Kongyuan;DONG Xiaoxin;Deng Linfeng(School of Mechanical and Electronic Engineering,Lanzhou University of Technology Lanzhou,730050,China;School of Intelligent Manufacturing and Electrical Engineering,Guangzhou Institute of Science and TechnologyGuangdong,510540,China)
出处
《振动.测试与诊断》
北大核心
2025年第6期1128-1135,1272,共9页
Journal of Vibration,Measurement & Diagnosis
基金
国家自然科学基金资助项目(62241308,51675253)。
关键词
深度学习
半监督学习
跨域故障诊断
辅助分类器
deep learning
semi-supervised learning
cross-domain fault diagnosis
auxiliary classifiers
