小样本下基于ProtoNet-AE的半监督跨工况故障诊断
Semi-supervised Cross-Condition Fault Diagnosis Based on ProtoNet-AE in Small-Sample Setting
梁城 1马萍 1王聪 1李新凯 1张宏立1
作者信息
- 1. 新疆大学 电气工程学院,乌鲁木齐 830017
- 折叠
摘要
针对小样本下跨工况的轴承故障诊断问题,提出一种结合了原型网络的深度自编码器(Prototype network-Autoencoder:ProtoNet-AE)半监督故障诊断框架.首先,构造基于时频图的小样本故障样本集,并建立无标签数据样本;然后,利用ProtoNet-AE中的AE对源域标记样本与无标签样本集进行半监督自适应特征提取和模型预训练,并设计了基于原型网络和自适应特征提取为一体的目标函数用于减小域分布差异;最后,通过少量目标域样本进行模型微调,提高了模型在目标域上的分类准确率和泛化性.通过跨工况下小样本故障诊断实验表明,对比于其他模型,所提模型均具有较强的可行性和有效性.
Abstract
A semi-supervised fault diagnosis framework,called ProtoNet-AE(Prototype network-Autoencoder),is proposed to address the cross-condition fault diagnosis problem in a small-sample setting for bearings.Firstly,a small sample set of fault samples based on time-frequency spectrograms is constructed,and unlabeled data samples are established.Next,an autoencoder(AE)is utilized for semi-supervised adaptive feature extraction on both the labeled samples from the source domain and the unlabeled dataset.Simultaneously,a target function is designed,integrating the prototype network and AE adaptive feature extraction,to perform model pre-training.Finally,the model is fine-tuned using a small number of labeled samples from the target domain to enhance its classification accuracy and generalization in the target domain.Experimental results in small-sample fault diagnosis under varying working conditions demonstrate the strong feasibility and effectiveness of the proposed framework when compared to other models.
关键词
小样本/原型网络/深度自编码器/跨工况故障诊断/半监督Key words
small-sample/prototype network/deep autoencoder/cross-condition fault diagnosis/semi-supervised引用本文复制引用
基金项目
国家自然科学基金(52065064)
新疆维吾尔自治区自然科学基金青年项目(2022D01C367)
出版年
2024
NETL
NSTL
万方数据