基于贝叶斯深度学习的外啮合齿轮泵健康状态识别研究  

Health Status Recognition of Gear Pump Based on Bayesian-LSTM

作　　者：郭锐[1,2,3] 王涛[1] 李永涛 王建伟 蔡伟 赵静一 GUO Rui;WANG Tao;LI Yongtao;WANG Jianwei;CAI Wei;ZHAO Jingyi(Hebei Key Laboratory of Heavy Machinery Fluid Power Transmission and Control,Yanshan University,Qinhuangdao 066004;State Key Laboratory of Fluid Power and Electromechanical System,Zhejiang University,Hangzhou 310027;Key Laboratory of Advanced Forging Technology and Science,Ministry of Education,Yanshan University,Qinhuangdao 066004;Hebei Key Laboratory of Specialized Transportation Equipment,Qinhuangdao 066004)

机构地区：[1]燕山大学河北省重型机械流体动力传输与控制重点实验室,秦皇岛066004 [2]浙江大学流体动力与机电系统国家重点实验室,杭州310027 [3]燕山大学先进锻压成形技术与科学教育部重点实验室,秦皇岛066004 [4]河北省特种运载装备重点实验室,秦皇岛066004

出　　处：《机械工程学报》2023年第16期397-405,共9页Journal of Mechanical Engineering

基　　金：国家自然科学基金(52075469,12173054);国家重点研发计划(2019YFB2005204);陕西省液压技术重点实验室开放基金(YYJS2022KF12)资助项目。

摘　　要：液压泵健康状态的精准识别是液压元件预防性维护的一部分。采用贝叶斯算法对长短时记忆网络(Long short-term memorg,LSTM)进行优化,构建一种基于贝叶斯长短时记忆网络(Bayesian-LSTM)的外啮合齿轮泵健康状态识别模型,完成了对外啮合齿轮泵健康状态的识别。首先运用改进的变分模态分解(Improved variational modal,IVMD)方法对蕴含丰富健康信息的样本泵振动信号进行分解重构,基于时域、频域及时频域对重构信号进行特征提取,将所选特征归一化处理后构造了能够表征振动信号的特征矩阵。最后将归纳好的特征矩阵与标签输入到Bayesian-LSTM(以下简称BDL)模型中进行训练,进而得到外啮合齿轮泵健康识别模型,为验证BDL模型在外啮合齿轮泵健康状态识别上的优越性,将其与LSTM模型及RNN模型进行对比。结果表明,BDL模型对样本泵的预测精度达到了98.3%以上,相比LSTM模型及RNN模型优势明显,可以为深度学习在液压元件状态识别的应用提供参考。Accurate recognition of the health status for hydraulic pumps is part of preventive maintenance(PM).A health status recognition model of external gear pumps based on Bayesian-Long short-term memorg(BDL)combines the Bayesian method and the deep learning algorithm to realize the health state recognition of the external gear pumps is proposed.Firstly,improved variational modal decomposition(IVMD)is used to decompose and reconstruct the vibration signal of the sample pump,whose features are extracted based on the time domain,frequency domain and time-frequency domain.After the selected features are normalized,the feature matrix representing the vibration signal is constructed.Finally,the feature matrix and labels are used to train the BDL model,and the external gear pump health status recognition model is obtained.In order to verify the superiority of the BDL in recognition for the health status of external gear pumps,a comparison with models of LSTM and RNN has carried.The results show that the prediction accuracy of BDL is more than 98.3%,which has obvious advantages than LSTM and RNN.The conclusion can provide a reference for the application of deep learning in status recognition of hydraulic components.

关 键 词：外啮合齿轮泵 健康状态识别 加速退化试验 贝叶斯长短时记忆网络 

分 类 号：TH325[机械工程—机械制造及自动化]