摘要
结合深度学习理论,将一维卷积神经网络运用于振动信号故障诊断,相较于传统方法,提取特征简单且高效。为进一步优化一维卷积结构,弥补其在信号所有位置的寻找模式,联系周期内的故障特征,提出一种新型DSCNN-GRU网络。该模型融合了深度可分离卷积的轻量快捷,降低了一维卷积结构参数;加入门控机制,可记忆分析故障点的信号特征,联系周期内的信号关系,更好地捕捉信号故障特征,提升对时间序列的敏感性。提出一种跟踪梯度优化Adam算法,解决模型随时间窗振荡问题。通过采集的减速机滚动轴承数据研究表明,该算法平均故障识别率可达94%以上,分类效果明显,泛化能力强。
Being based on the theory of deep learning,one-dimensional Convolutional Neural Networks is applied to fault diagnosis of bearing vibration signals.Compared with traditional methods,extraction feature is simple and efficient.In order to further optimize the one-dimensional convolution structure,smooth over its search mode at all positions of the signal,and connect the fault characteristics in the period,a new DSCNN-GRU network model is proposed.The model combines the lightweight and fast of Depthwise Separable Convolution,and reduces the structural parameters of one-dimensional convolution.By adding gating mechanism,the signal characteristics of fault points can be memorized and analyzed,and the signal relationship in the period can be linked to better capture the signal fault characteristics and enhance the sensitivity of temporal series.An Adam algorithm for tracking gradient optimization is proposed to solve the problem of model with time windows oscillation.The data collected from the reducer rolling bearing shows that the average fault recognition rate of the algorithm can reach more than 94%,the classification effect is obvious,and the generalization ability is stronger.
作者
汪洋
郭利进
Wang Yang;Guo Lijin(School of Electrical Engineering and Automation,Tianjin Polytechnic University,Tianjin 300387,China)
出处
《机械科学与技术》
CSCD
北大核心
2020年第2期258-266,共9页
Mechanical Science and Technology for Aerospace Engineering
基金
天津市自然科学基金青年项目(16JCQNJC03800)资助.
关键词
卷积神经网络
深度可分离卷积
门控机制
故障诊断
滚动轴承
convolutional neural networks
depthwise separable convolution
gate mechanism
fault diagnosis
rolling bearing
model
algorithm
fault recognition
作者简介
汪洋(1993-),硕士研究生,研究方向为过程监测与故障诊断,wangyangclover@163.com;通信作者:郭利进,教授,博士,Doctor-guo@tjpu.edu.cn
