跳连接变分自编码器与CNN相结合的滚动轴承故障诊断方法
Fault Diagnosis Method of Rolling Bearing Combining Jump Connected Variational Auto-encoder with CNN
张洪亮 1余其源 1王锐2
作者信息
- 1. 安徽工业大学 管理科学与工程学院,安徽马鞍山 243032
- 2. 苏州大学 轨道交通学院,江苏苏州 215131
- 折叠
摘要
针对滚动轴承故障率小、不易收集故障数据的问题,提出基于跳跃连接变分自编码器与宽核深度卷积神经网络相结合的小样本故障诊断方法.该方法首先在变分自编码器的编码和解码之间引入跳跃连接结构,并将Tanh作为网络的激活函数,进而提高生成样本的特征多样性;其次,构建宽核深度卷积网络诊断模型,该模型可以提高从振动信号中提取故障特征的能力;最后,经生成样本扩充的数据集作为模型输入,提高训练集包含的特征信息量,实现小样本下的故障诊断.实验分析表明,所提方法在小样本情形下能生成有效的伪样本并具有较高的诊断精度.
Abstract
For the problem that the failure rate of rolling bearings is small and it is not easy to collect fault data,a novel rolling bearing fault diagnosis method with small samples is proposed,which combines respective advantages of jumping connection variational auto-encoder and deep convolution neural network with wide kernel.The proposed method firstly introduces a jump connection structure between encoding and decoding of the variational auto-encoder,and Tanh is used as the activation function of the network,thus improving the feature diversity of the generated samples.Secondly,the diagnosis model of wide kernel deep convolution network is constructed,aiming to enhance the capability of fault feature extraction from vibration signals.Finally,the data set expanded by the generated samples is used as the model input to improve the amount of feature information contained in the training set,thereby realizing bearing fault diagnosis under small samples.Experimental analysis shows that the proposed method can generate effective fake samples and gains high diagnostic accuracy in the case of small samples.
关键词
故障诊断/跳跃连接变分自编码器/数据生成/宽核深度卷积神经网络Key words
fault diagnosis/jump connected variational auto-encoder/data generation/deep convolution neural network with wide kernel引用本文复制引用
基金项目
安徽省自然科学基金(2108085MG236)
安徽省普通高校重点实验室开放基金重点项目(CS2021-ZD01)
出版年
2024
国家科技期刊平台
NSTL
维普
万方数据