小样本下基于原型网络的轴向柱塞泵故障诊断模型  

Fault diagnosis modal of axial piston pump based on prototypical network with small sample size

作　　者：范佳祺 兰媛 黄家海 熊晓燕 李国彦 李利娜 FAN Jia-qi;LAN Yuan;HUANG Jia-hai;XIONG Xiao-yan;LI Guo-yan;LI Li-na(School of Mechanical and Transportation Engineering,Taiyuan University of Technology,Taiyuan 030024,China;Key Laboratory of New Sensors and Intelligent Control of Ministry of Education,Taiyuan University of Technology,Taiyuan 030024,China)

机构地区：[1]太原理工大学机械与运载工程学院,山西太原030024 [2]太原理工大学新型传感器与智能控制教育部重点实验室,山西太原030024

出　　处：《机电工程》2023年第4期584-591,共8页Journal of Mechanical & Electrical Engineering

基　　金：山西省关键核心技术和共性技术研发攻关项目(2020XXX001);山西省应用基础研究计划青年科技基金项目(202103021223090)。

摘　　要：在实际工程应用中,有限的故障样本数量及噪声都影响轴向柱塞泵故障诊断的效果,所以,如何提高模型在小样本、噪声条件下轴向柱塞泵故障诊断的性能是一个亟待解决的问题。在样本数量有限、噪声条件下,采用基于深度学习的故障诊断方法会出现过拟合、诊断准确率下降的问题,为此,提出了一种小样本条件下基于原型网络的轴向柱塞泵故障诊断模型(方法)。首先,搭建了轴向柱塞泵故障诊断模型,并等量随机抽取了每个故障的样本以构建多个任务,模型使用一维卷积神经网络作为主干,每个任务中包含当前模型、支持集、查询集;然后,利用模型将样本映射到特征空间,在特征空间中,模型使用支持集的同类样本构建了原型点,并逐个将查询集样本与多个原型点进行了距离度量,实现了轴向柱塞泵不同故障的分类;最后,为了验证基于原型网络的轴向柱塞泵故障诊断模型的有效性,采集了轴向柱塞泵不同元件发生故障时产生的振动信号,并使用上述诊断模型对此进行了故障识别实验;为了验证该诊断模型的优越性,将其与基于卷积神经网络等的模型进行了性能对比。实验结果表明:在样本有限的条件下,采用基于原型网络的轴向柱塞泵故障诊断模型的准确率达到85%以上;同时,在噪声条件下,采用基于原型网络的模型的准确率也能达到85%以上。研究结果表明:基于原型网络的模型的诊断性能优于卷积神经网络模型与传统方法。In practical engineering application,the limited number of fault samples and noise both affect the effect of fault diagnosis of axial piston pump.Therefore,how to improve the performance of fault diagnosis of axial piston pump under the condition of small samples and noise is an urgent problem to be solved.Aiming at the problem that fault diagnosis model based on deep learning tended to be over-fitted,and the performance of the model was greatly affected under the condition of limited number of samples and noise,a fault diagnosis model of axial piston pump based on prototypical network with small sample size was proposed.Firstly,a fault diagnosis model of axial piston pump was established,and equal samples of each fault were randomly selected to construct multiple tasks.One-dimensional convolutional neural network was used as the backbone of the model,and each task contains current model,support set and query set.Then,the samples were mapped to the feature space by using the model.In the feature space,the prototype points were constructed by using the similar samples of the support set,and the distance between the sample of the query set and several prototype points was measured one by one,thus realizing the classification of different faults of the axial piston pump.Finally,in order to verify the validity of the fault diagnosis model of axial piston pump based on prototype network,the vibration signals generated by different components of axial piston pump were collected,and the fault identification experiment was carried out by using the above diagnosis model.In order to verify the superiority of this diagnostic model,its performance was compared with that of the model based on convolutional neural network.The experiment results show that the accuracy of the axial piston pump fault diagnosis model based on prototypical network is more than 85%under the condition of limited samples.At the same time,the accuracy under the condition of noise is also more than 85%.The research results show that the performance of m

关 键 词：容积泵 深度学习 有限样本数量 抗噪性 故障分类识别 诊断准确率 结果可视化 

分 类 号：TH322[机械工程—机械制造及自动化] TH17