• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2019, Vol. 55 ›› Issue (7): 58-64.doi: 10.3901/JME.2019.07.058

• 基于深度学习的机械装备故障预测与健康管理 • 上一篇    下一篇

基于萤火虫优化的核自动编码器在中介轴承故障诊断中的应用

王奉涛, 刘晓飞, 敦泊森, 邓刚, 韩清凯, 李宏坤   

  1. 大连理工大学振动工程研究所 大连 116024
  • 收稿日期:2018-05-21 修回日期:2018-11-02 出版日期:2019-04-05 发布日期:2019-04-05
  • 通讯作者: 李宏坤(通信作者),男,1974年出生,博士,教授,博士研究生导师。主要研究方向为旋转机械信号处理与故障诊断,振动与噪声。E-mail:lihk@dlut.edu.cn
  • 作者简介:王奉涛,男,1974年出生,博士,副教授。主要研究方向为机械系统信号处理、故障诊断与寿命预测。E-mail:wangft@dlut.edu.cn
  • 基金资助:
    国家自然科学基金资助项目(51875075,51375067)。

Application of Kernel Auto-encoder Based on Firefly Optimization in Intershaft Bearing Fault Diagnosis

WANG Fengtao, LIU Xiaofei, DUN Bosen, DENG Gang, HAN Qingkai, LI Hongkun   

  1. Institute of Vibration Engineering, Dalian University of Technology, Dalian 116024
  • Received:2018-05-21 Revised:2018-11-02 Online:2019-04-05 Published:2019-04-05

摘要: 随着科技进步与工业规模的快速壮大,现代工业监测领域步入大数据时代,如何自动地从大规模原始信号中提取故障特征并诊断是一个重要课题。为了提高深度自动编码网络处理非线性问题的能力,提出一种基于核函数与去噪自动编码器(Denosing auto-encoder,DAE)的深度神经网络方法。采用径向基核函数改进传统的去噪自动编码器,提出核去噪自动编码器(Kernel denosing auto-encoder,KDAE);构建包含一个KDAE层与多个AE层的深度神经网络对故障特征进行层层提取,并添加softmax分类层;采用误差反向传播算法对网络参数进行微调,并采用混沌萤火虫算法优化核参数与深度网络中的待定参数,得到故障诊断模型。针对传统自动编码器泛化性较差的问题,在目标函数中添加L2惩罚项。通过航空发动机中介轴承典型故障试验数据,验证了所提方法与传统去噪自动编码网络相比能够得到更高的准确率。

关键词: 高斯核函数, 故障诊断, 深度学习, 中介轴承, 自动编码器

Abstract: With the rapid development of scientific technological progress and industrial scale, modern industrial monitoring field has entered the era of big data. It is an important task to automatically extract fault features from large scale raw vibration data and make fault diagnosis. In order to further improve the ability of the deep auto-encoder network to deal with the nonlinear problem, a deep neural network method based on kernel function and denoising auto-encoder (DAE) is proposed. The traditional denoising auto-encoder is improved by radial basis kernel function, and the kernel denoising auto-encoder (KDAE) is proposed. A deep neural network consisting of one KDAE layer and multiple AE layers is constructed to extract the fault features, and the softmax classification layer is added as classifier layer. The error back propagation algorithm is used to fine-tune the network parameters, and chaos firefly algorithm is used to optimize the undetermined parameters of the kernel parameters, then the fault diagnosis model is obtained. In response to the problem of poor generalization of traditional auto-encoder, L2 penalty items are added to the target function. It is verified that the proposed method is more accurate than the traditional denoising auto-encoder network through the typical failure test data of aero-engine intermediate bearing.

Key words: auto-encoder, deep learning, fault diagnosis, Gaussian kernel function, intershaft bearing

中图分类号: