航空学报 > 2022, Vol. 43 Issue (8): 625201-625201   doi: 10.7527/S1000-6893.2021.25201

故障诊断技术在航空航天领域中的应用专栏

深度残差对冲网络及其在滚动轴承故障诊断中的应用

康玉祥1, 陈果1, 尉询楷2, 周磊2   

  1. 1. 南京航空航天大学 民航学院, 南京 210016;
    2. 北京航空工程技术研究中心, 北京 100076
  • 收稿日期:2021-01-05 修回日期:2021-07-21 出版日期:2022-08-15 发布日期:2021-07-20
  • 通讯作者: 陈果,E-mail:cgzyx@263.net E-mail:cgzyx@263.net
  • 基金资助:
    国家科技重大专项(J2019-IV-004-0071)

Deep residual hedging network and its application in fault diagnosis of rolling bearings

KANG Yuxiang1, CHEN Guo1, WEI Xunkai2, ZHOU Lei2   

  1. 1. College of Civil Aviation, Nanjing University of Aeronautics and Astronautics, Nanjing 210016, China;
    2. Beijing Aeronautical Engineering Technical Research Center, Beijing 100076, China
  • Received:2021-01-05 Revised:2021-07-21 Online:2022-08-15 Published:2021-07-20
  • Supported by:
    National Science and Technology Major Project (J2019-IV-004-0071)

摘要: 提出了一种新的深度残差对冲网络模型。首先,该模型借助Inception堆叠思想提出了堆叠卷积对冲结构块以加快网络收敛速度;然后,设计了新的恒等映射块,实现了输入层与中间各层的残差连接;最后,在全连接层引入Squash函数,防止损失梯度的发散。将提出的深度残差对冲网络应用于滚动轴承故障诊断,在预处理中将滚动轴承的振动加速度时域信号通过快速傅里叶变换得到的频谱图直接作为网络输入,从而简化了数据的预处理工作。利用两组实际的滚动轴承故障数据进行方法验证,并与18层深度残差网络(Resnet18)、卷积神经网络(CNN)等其他方法进行了对比验证。结果表明所提深度残差对冲网络模型的测试精度较其他模型高约2%,且训练时间能缩短1/3,充分表明本文方法具有很强的鲁棒性和收敛速度快等优点。

关键词: 深度学习, 残差网络, 对冲结构, Squash函数, 滚动轴承, 故障诊断

Abstract: A new depth residual hedging network model is proposed, in which, with the help of Inception stacking idea, a stacked convolution hedging structural block was proposed to accelerate the convergence speed of the network, and a new identity mapping block is designed to realize the residual connection between the input layer and the middle layer. Moreover, the follow-up Squash function is introduced in the full connection layer to prevent the divergence of loss gradient. The depth residual offset network is applied to the fault diagnosis of the rolling bearing. In the preprocessing, the frequency spectrum of the vibration acceleration signal of the rolling bearing is directly taken as the input of the network, thus, simplifying the preprocessing of the data. Finally, two sets of actual rolling bearing failure data are used for methods validation, and to be compared with 18 Layer Residual Networks Deep Residual Networks (Resnet18), Convolutional Neural Networks (CNN) and other verification methods. The results show that of the depth of residual hedge network test accuracy of the model proposed is estimated at 2% more than other models, and training time can be shortened to one third, which fully suggests that the method has strong robustness and fast convergence rate.

Key words: deep learning, residual network, hedge structure, Squash functions, rolling bearing, fault diagnosis

中图分类号: