基于局部描述子的小样本轴承故障诊断方法  

Few-shot Learning Fault Diagnosis Method Based on Local Descriptor

作　　者：赵志宏[1,2] 陶旭 武超 ZHAO Zhihong;TAO Xu;WU Chao(State Key Laboratory of Mechanical Behavior and System Safety of Traffic Engineering Structures,Shijiazhuang Tiedao University,Shijiazhuang 050043,China;School of Information Science and Technology,Shijiazhuang Tiedao University,Shijiazhuang 050043,China)

机构地区：[1]石家庄铁道大学省部共建交通工程结构力学行为与系统安全国家重点实验室,河北石家庄050043 [2]石家庄铁道大学信息科学与技术学院,河北石家庄050043

出　　处：《铁道车辆》2024年第5期23-29,共7页Rolling Stock

基　　金：国家自然科学基金(11972236,12172234)。

摘　　要：深度学习由于其在特征表示方面的优势,近年来已被应用于很多领域,但是在故障诊断领域很难获取大量的故障样本来训练模型。针对这一问题,提出了一种基于局部描述子的小样本轴承故障诊断方法,利用改进的深度最近邻神经网络(Deep Nearest Neighbor Neural Network,DN4)进行故障诊断。首先通过短时傅里叶变换将振动信号转换为二维时频图像,特征提取主干网络采用ResNet-12网络,引入高效多尺度注意力机制(Efficient Multi-scale Attention,EMA)来更好地提取故障特征,得到轴承故障的局部描述子,最后利用K近邻算法来得到故障类别。为验证所提方法的有效性,进行了不同样本数量、跨工况条件下的轴承故障诊断实验,实验结果表明,文章所提方法在小样本条件下具有较好的故障识别效果和泛化性能,具有一定的工程应用价值。Due to its advantages in feature description,deep learning has been applied to many fields in recent years.However,in the field of fault diagnosis,it is difficult to obtain a large number of fault samples to train models.To solve this problem,this paper proposes a bearing fault diagnosis method based on local descriptors according to localized feature descriptions,and uses the improved Deep Nearest Neighbor Neural Network(DN4)for fault diagnosis.Firstly,the vibration signal is transformed into a two-dimensional time-frequency image using a short-time Fourier transform.The ResNet-12 network is employed to extract the backbone network features,and Efficient Multi-scale Attention(EMA)is implemented to enhance the extraction of fault features.The local descriptor of bearing faults is obtained.Finally,the fault category is obtained using the K-nearest neighbor algorithm.To ascertain the efficacy of the proposed methodology,tests of bearing fault diagnosis were conducted with varying numbers of samples and encompassing diverse working conditions.The results of the test demonstrate that the proposed method exhibits a favorable effect on fault identification and generalization performance based on few-shot learning,as well as possessing certain engineering application value.

关 键 词：轴承 小样本 故障诊断 局部描述子 高效多尺度注意力机制 

分 类 号：U270.331[机械工程—车辆工程]