基于HTMFDE以及ICN...的滚动轴承寿命状态识别方法_董绍江.pdf

第 20 卷 第 2 期2023 年 2 月铁道科学与工程学报Journal of Railway Science and EngineeringVolume 20 Number 2February 2023基于HTMFDE以及ICNN的滚动轴承寿命状态识别方法董绍江1，刘文龙1，方能炜2，胡小林2，余腾伟1(1.重庆交通大学 机电与车辆工程学院，重庆 400074；2.重庆工业大数据创新中心有限公司，重庆 400707)摘要：针对滚动轴承退化性能难以评估、寿命状态难以识别的难题，提出一种结合层次时移多尺度波动散布熵(Hierarchical Time-shifted Multiscale Fluctuation Dispersion Entropy，HTMFDE)、JRD 距离(Jensen-Renyi divergence，JRD)以及改进卷积神经网络(Improved convolution neural network，ICNN)的轴承寿命状态识别新方法。首先，在传统多尺度波动散布熵的基础上，将传统均值粗粒化过程替换为改进的时移粗粒化过程，以解决传统均值粗粒化导致信号幅值特征丢失的问题。同时引入层次分解理论，克服多尺度分析方法不能全面提取不同频段故障特征的不足，得到最终的HTMFDE。之后利用HTMFDE方法提取滚动轴承信号的多维状态特征量，并进行归一化形成一组概率分布，计算轴承正常信号与故障信号之间的JRD距离作为性能退化指标。其次，根据构建的JRD性能退化曲线，划分轴承寿命状态并制作数据集，通过标签化的样本训练具有双层多尺度特征提取层的卷积神经网络，建立滚动轴承寿命状态识别模型。为了加快模型的收敛速度，对每层卷积进行批量归一化操作，同时采用全局池化代替全连接层以提升模型的训练效率。最后，在2组不同的轴承数据集上进行实验。实验结果表明，根据HTMFDE构建的JRD性能退化曲线能够精准地识别轴承性能退化起始点以及刻画轴承的退化趋势，所提出的ICNN模型在SNR=010 dB环境中平均识别正确率为98.5%，能够准确地识别轴承寿命状态，验证了所提方法的实用性以及有效性。关键词：寿命状态识别；滚动轴承；层次时移多尺度波动散布熵；JRD距离；改进卷积神经网络中图分类号：TH133.33 文献标志码：A 开放科学(资源服务)标识码(OSID)文章编号：1672-7029（2023）02-0723-12Recognition method of rolling bearing life state based on HTMFDE and ICNNDONG Shaojiang1,LIU Wenlong1,FANG Nengwei2,HU Xiaolin2,YU Tengwei1(1.School of Mechantronics and Vehicle Engineering,Chongqing Jiaotong University,Chongqing 400074,China;2.Chongqing Industrial Big Data Innovation Center Co.,Ltd.,Chongqing 400707,China)Abstract:Aiming at the problems that the degradation performance of rolling bearings is difficult to evaluate and the life state is difficult to identify,a new method for the identification of bearing life state combining hierarchical 收稿日期：2022-03-08基金项目：国家自然科学基金资助项目(51775072)；重庆市科技创新领军人才支持计划项目(CSTCCCXLJRC201920)；重庆市高校创新研究群体项目(CXQT20019)；重庆市北碚区科学技术局技术创新与应用示范项目(2020-6)；城市轨道交通车辆系统集成与控制实验室开放基金资助项目(CKLURTSIC-KFKT-202007)通信作者：董绍江(1982)，男，山东烟台人，教授，博士，从事旋转机械系统状态分析和故障诊断、趋势预测和大数据挖掘研究；Email：DOI:10.19713/ki.43-1423/u.T20220421铁 道 科 学 与 工 程 学 报2023 年 2月time-shifted multiscale fluctuation dispersion entropy(HTMFDE),jensen-renyi divergence(JRD)and improved convolutional neural network(ICNN)was proposed.First,on the basis of the traditional multi-scale fluctuation dispersion entropy,the traditional mean coarse-graining process was replaced with an improved time-shifted coarse-graining process to solve the problem of loss of signal amplitude characteristics caused by traditional mean coarse-graining.Meanwhile,the hierarchical decomposition theory was introduced to overcome the deficiency that the multi-scale analysis method cannot fully extract the fault features of different frequency bands,thereby the final HTMFDE was obtained.Then,the HTMFDE method was used to extract the multi-dimensional state feature quantities of the rolling bearing signals,which were then normalized to form a set of probability distributions,and the jensen-renyi divergence between the normal and faulty signals of the bearing was calculated as the performance degradation index.Second,according to the constructed jensen-renyi divergence performance degradation curve,the bearing life state was divided and a data set was produced,and a convolutional neural network with a double-layer multi-scale feature extraction layer was trained through the labeled samples to establish a rolling bearing life state recognition model.In order to speed up the convergence of the model,batch normalization was performed on each layer of convolution,and global pooling was used to replace the fully connected layer to improve the training efficiency of the model.Finally,experiments were conducted on two different bearing datasets.The experimental results show that the jensen-renyi divergence performance degradation curve constructed according to HTMEDE can accurately identify the starting point of bearing performance degradation and describe the degradation trend of the bearing.The proposed ICNN model has an average recognition accuracy of 98.5%in the SNR=010 dB environment,and can accurately identify the bearing life state.This verified the practicability and effectiveness of the proposed method.Key words:life state identification;rolling bearing;hierarchical time-shifted multiscale fluctuation dispersion entropy;Jensen-Renyi divergence;improved convolutional neural network 滚动轴承作为轨道、航空、风电等领域中不可缺少的零部件，同时也是最容易发生故障的零部件之一，其工作状态将直接影响机械设备的运行状态1，若不能及时监测其是否出现故障并采取一定应对措施，一旦轴承失效，可能造成严重的事故和大量的财产损失。因此，研究有效的滚动轴承寿命状态识别方法具有十分重要的意义。轴承寿命状态识别有2个关键问题需要解决：1)构建合适的轴承性能退化评估指标；2)建立有效的轴承寿命状态识别模型2。传统描述轴承退化趋势的方法主要是提取时域、频域和时频域的特征量来描述，如利用峭度和均方根值等。但滚动轴承的振动信号受运行环境、工况等因素影响，常常表现出非线性特征，仅仅依靠传统单一时频域的特征量已无法精准地描述轴承性能退化趋势。周建民等3提出基于特征优选结合优化支持向量机的性能评估方法，较好地刻画了轴承的性能退化趋势，但需要同时计算10余种时域特征参数，计算过程复杂繁琐。董绍江等4通过构造特征噪声能量比(Feature-to-Noise Energy Ratio，FNER)也较好地描述了轴承退化趋势，但需要提前明确轴承的故障频率，不适用于实际工况下未知轴承的性能评估。近年来，基于熵理论的非线性特征量在故障特征提取领域得到了广泛应用，如多尺度样本熵(Multiscale Sample Entropy，MSE)5、多 尺 度 排 列 熵(Multiscale Permutation Entropy，MPE)6和多尺度散布熵(Multi-scale Dispersion Entropy，MDE)7等。JIAO等8将MSE与能量矩阵相结合，从多个轴承信号中获取故障特征，结合最小二乘支持向量机准确地识别出了轴承的故障程度。LI等9提出一种多尺度符号模糊熵(Multiscale Symbolic Fuzzy Entropy，MCFE)并用于机械故障的提取。然而，上述方法都有其局限性，如