上海交通大学学报 ›› 2022, Vol. 56 ›› Issue (2): 182-190.doi: 10.16183/j.cnki.jsjtu.2021.001

• • 上一篇    下一篇

基于特征金字塔卷积循环神经网络的故障诊断方法

刘秀丽(), 徐小力   

  1. 北京信息科技大学 现代测控技术教育部重点实验室,北京 100192
  • 收稿日期:2021-01-04 出版日期:2022-02-28 发布日期:2022-03-03
  • 作者简介:刘秀丽(1986-),女,山东省济南市人,博士,助理研究员,现主要从事机电装备故障诊断与预测研究.E-mail: liuxiulilw@163.com.
  • 基金资助:
    国家自然科学基金项目(51975058);国家重点研发计划(2020YFB1713203)

A Fault Diagnosis Method Based on Feature Pyramid CRNN Network

LIU Xiuli(), XU Xiaoli   

  1. Key Laboratory of Modern Measurement and Control Technology of the Ministry of Education, Beijing Information Science and Technology University, Beijing 100192, China
  • Received:2021-01-04 Online:2022-02-28 Published:2022-03-03

摘要:

变工况、变载荷设备部件不同故障的特征在信号中所占比例和位置不固定,且包括大量不同场景下的原始振动信号的多尺度复杂性.对此,提出一种基于特征金字塔网络(FPN)的卷积循环神经网络(CRNN)滚动轴承故障诊断方法.利用卷积神经网络(CNN)框架,并联CNN的卷积层和循环神经网络(RNN)中的长短时记忆(LSTM)层,形成新的CRNN,以充分利用CNN对空间域信息和RNN对时域信息的学习能力;在每一层中权值共享,减少网络参数;利用FPN构建全新特征图,输入一维信号和堆叠后形成的二维信号,对传感器采集的信号进行特征提取,实现故障诊断.利用行星齿轮箱进行故障试验,并进行5折交叉验证,该方法的诊断准确率平均值为99.20%,比基本神经网络模型至少高3.62%,表明该方法诊断精度高、鲁棒性强;利用凯斯西储大学轴承数据集进行验证,证明该方法具有良好的泛用性;利用t-SNE方法对模型的特征学习效果进行可视化分析,结果表明不同故障类别特征具有良好的聚类效果.

关键词: 卷积循环神经网络, 特征金字塔, 故障诊断, 特征可视化

Abstract:

Aimed at the problems that the proportion and position of different fault characteristics of equipment components under variable working conditions and variable load in the signal are not fixed, and include the multi-scale complexity of the original vibration signal in a large number of different scenarios, a convolutional recurrent neural network (CRNN) rolling bearing fault diagnosis method based on feature pyramid network (FPN) was proposed. Using the convolution neural network (CNN) framework, the convolution layer of CNN and the long and short-term memory (LSTM) layer of recurrent neural network (RNN) were connected in parallel to form a new CRNN, so as to make full use of the learning ability of CNN to spatial domain information and RNN to time domain information. The weights were shared in each layer to reduce network parameters. A novel feature map was constructed using FPN, and one-dimensional signal and two-dimensional signal formed after stacking were input to extract the feature of the signal collected by the sensor, and realize fault diagnosis. The average diagnostic accuracy of this method is 99.20%, which is at least 3.62% higher than that of the basic neural network model, indicating that this method has a high diagnostic accuracy and a strong robustness. Using the bearing data set of Case Western Reserve University, it is proved that the method has a good universality. The t-SNE method was used to visually analyze the feature learning effect of the model. The results show that different fault category features have good clustering effect.

Key words: convolutional recurrent neural network (CRNN), feature pyramid, fault diagnosis, feature visualization

中图分类号: