液压泵和滚动轴承多种样本量的改进多任务故障诊断  

Improved multi-task fault diagnosis of hydraulic pump and rolling bearingwith multiple sample sizes

作　　者：郑直 曾魁魁 何玉灵[2] 李克 王志军 ZHENG Zhi;ZENG Kuikui;HE Yuling;LI Ke;WANG Zhijun(College of Mechanical Engineering,North China University of Science and Technology,Tangshan 063210,China;Hebei Key Laboratory of Electric Machinery Health Maintenance and Failure Prevention,Department of Mechanical Engineering,North China Electric Power University,Baoding 071000,China)

机构地区：[1]华北理工大学机械工程学院,河北唐山063210 [2]华北电力大学电机健康维护与故障预防河北省重点实验室,河北保定071000

出　　处：《振动与冲击》2024年第4期270-278,共9页Journal of Vibration and Shock

基　　金：河北省自然科学基金资助项目(E2022209086);国家自然科学基金资助项目(52177042;51777074);唐山市科技创新团队培养计划项目(21130208D);河北省科技重大专项项目(22282203Z)。

摘　　要：基于充足样本的多个设备元件导致多任务学习网络规模庞大,轻微和严重的跨元件零样本问题难度大。在多种样本量(充足样本和零样本)下,针对基于充足故障样本的多元件诊断网络规模过于庞大问题,引入MicroNet方法对多任务学习网络进行轻量化处理,然后利用热重启余弦退火算法优化上述网络,提出一种多任务轻量化学习网络模型,改善多任务学习网络的准确率和效率。针对更高难度的跨元件零样本问题,引入元学习方法进一步改进上述MT-MN-CA,进而提出一种改进多任务轻量化学习网络模型,解决轻微和严重的跨元件零样本问题。通过实测液压泵和滚动轴承故障验证所提两个网络模型的有效性和优越性,试验结果表明所提网络具有很高的实时性和准确率。The multiple equipment components based on sufficient samples result in a large scale of multi-task learning network,and the slight and serious cross-component zero-sample problem has not been studied and is difficult.Under multiple sample sizes(sufficient samples and zero-sample),aiming at the problem that the scale of multi-component diagnosis network based on sufficient fault samples is too large,the MicroNet method was introduced to lighten the MTL network,and then the network was optimized by cosine annealing with warm restart algorithm,the lightweight MTL network model is proposed,thus the accuracy and efficiency of MTL network were improved.Aiming at the more difficult cross-component zero-sample problem,the meta-learning method was introduced to further improve the above MT-MN-CA,and then the improved lightweight MTL network model was proposed to solve the slight and serious cross-component zero-sample problem.The effectiveness and superiority of the proposed two network models were verified by the measured faults of hydraulic pump and rolling bearing,the experimental results show that the proposed network has high real-time performance and accuracy.

关 键 词：多任务学习 轻量化 元学习 零样本 故障诊断 

分 类 号：TH137[机械工程—机械制造及自动化] TP277[自动化与计算机技术—检测技术与自动化装置]