基于改进多尺度卷积循环神经网络的滚动轴承故障研究  

Rolling bearing fault study based on an improved multi-scale convolutional recurrent neural network

作　　者：董绍江[1] 黄翔 夏宗佑 邹松 DONG Shaojiang;HUANG Xiang;XIA Zongyou;ZOU Song(School of Electromechanical and Vehicle Engineering,Chongqing Jiaotong University,Chongqing 400074,China)

机构地区：[1]重庆交通大学机电与车辆工程学院,重庆400074

出　　处：《振动与冲击》2024年第20期94-105,共12页Journal of Vibration and Shock

基　　金：重庆市教委科学技术研究项目(KJZD-K202300711);重庆市科技创新领军人才支持计划项目(CSTCCCXLJRC201920);重庆市高校创新研究群体(CXQT20019)。

摘　　要：针对传统卷积神经网络故障诊断方法提取特征不丰富,容易丢失故障敏感信息,且在单一尺度处理方法限制实际复杂工况下故障特性的深度挖掘问题,提出了注意力机制的多尺度卷积神经网络和双向长短期记忆(bi-directional long short-term memory,BiLSTM)网络融合的迁移学习故障诊断方法。该方法首先应用不同尺寸池化层和卷积核捕获振动信号的多尺度特征;然后引入多头自注意力机制自动地给予特征序列中的不同部分不同的权重,进一步加强特征表示的能力;其次利用BiLSTM结构引入双向性质提取特征前后之间的内部关系实现信息的逐层传递;最后利用多核最大均值差异减小源域和目标域在预训练模型中各层上的概率分布差异并利用少量标记的目标域数据再对模型进行训练。试验结果表明,所提方法在江南大学(JNU)、德国帕德博恩大学(PU)公开轴承数据集上平均准确率分别为98.43%和97.66%,该方法在重庆长江轴承股份有限公司自制的轴承故障数据集上也表现出了极高的准确率和较快的收敛速度,为有效诊断振动旋转部件故障提供了实际依据。A novel fault diagnosis method was proposed,which combines a multi-scale convolutional neural network with a bi-directional long short-term memory(BiLSTM)network using the attention mechanism.This approach addresses the issue of feature extraction in traditional fault diagnosis methods,which often result in limited representation of fault information and the inability to deeply explore fault characteristics under complex working conditions.Firstly,the method employed pooling layers and convolutional kernels of different sizes to capture multi-scale features from vibration signals.Then,a multi-head self-attention mechanism was introduced to automatically assign different weights to different parts of the feature sequence,further enhancing the ability to represent features.Additionally,the BiLSTM structure was used to extract the internal relationships between features before and after,enabling the progressive transmission of information.Finally,the maximum-kernel mean discrepancy was utilized to reduce the distribution differences between the source and target domains at various layers of the pre-trained model,and a small amount of labeled target domain data was used to further train the model.The experimental results show that the proposed method has an average accuracy of 98.43%and 97.66%on the JNU and PU open bearing datasets,respectively,and the method also shows a very high accuracy and fast convergence speed on the bearing fault dataset made by Chongqing Changjiang Bearing Co.and provides a practical basis for the effective diagnosis of vibration rotating component faults.

关 键 词：故障诊断 多尺度卷积神经网络 双向长短期记忆(BiLSTM)网络 多头自注意力 多核最大均值差异 

分 类 号：TH133.33[机械工程—机械制造及自动化] TH165.3