首页|改进的残差网络及其在滚动轴承早期故障诊断中的应用

改进的残差网络及其在滚动轴承早期故障诊断中的应用

扫码查看
原文链接
  • NETL
  • NSTL
  • 万方数据
基于残差网络的滚动轴承故障诊断已经取得一定的成果,但传统的残差网络只能将输入信号进行自下而上的单向特征提取,如果当前层丢失信号中的有用信息,后续层将无法弥补丢失的信息.特别是滚动轴承发生早期微弱故障时,故障特征容易被噪声所掩盖.如何利用残差网络,充分提取滚动轴承早期故障特征,是一个亟待解决的问题.为此,本文提出一种具有密集连接机制(dense connection residual network,DRN)的新型残差网络.在DRN中,每个隐藏层都与输入信号建立有向连接,再利用通道级联算法,将输入信号和每个隐藏层特征进行重构,从而修复深层模型中遗漏的有用信息,获得更完整的故障特征.在XJTU-SY数据集上进行实验,当信噪比达到0dB、-1dB、-2dB、-3 dB、-4dB时,DRN的准确率均保持在95％以上,说明该方法具有较好的鲁棒性.
Early Fault Diagnosis of Rolling Bearings Based on Improved Residual Network
In recent years,bearing fault diagnosis with residual networks has achieved certain results.But traditional residual networks can only perform bottom-up one-way feature extraction.If the current layer losses useful information in the signal,the subsequent layers can't compensate for the lost information.Especially in early fault diagnosis of bearings,fault features are easily masked by noise.So how to use residual networks to fully extract early fault features is an urgent problem.This article proposes a new residual network with dense connection mechanism(DRN).In DRN,each hidden layer establishes a directed connection with the input data,and through channel cascade algorithms,the features are reconstructed with the input data and hidden layers.Thereby DRN can repair the lost information and obtain more complete fault features.Experiments are conducted on the XJTU-SY datasets,and the effectiveness of DRN is demonstrated through ablation experi-ments.Compared with traditional methods,when the signal-to-noise ratio reaches 0dB,-1dB,-2dB,-3dB,-4dB,the accuracies of DRN keep over 95％.This indicates that the method has good feature extraction capabilities.

residual networkdense connection mechanismchannel cascade aggregationfeature repairearly fault diagnosis of bearing

陶洁、尹石磊、吴小明、赵志磊、邱海文

展开 >

湖南科技大学计算机科学与工程学院,湘潭 411201

湘潭市中心医院,湘潭 411100

残差网络 密集连接机制 通道级联算法 特征重构 轴承早期故障诊断

湖南省自然科学基金资助项目湖南省自然科学基金资助项目湖南省自然科学基金资助项目湖南省教育厅科研资助项目

2023JJ302652022JJ900032022JJ9001222C0262

2024

湖南工程学院学报(自然科学版)
湖南工程学院

湖南工程学院学报(自然科学版)

影响因子：0.265
ISSN：1671-119X
年,卷(期)：2024.34(2)