电工技术学报  2023, Vol. 38 Issue (10): 2675-2685    DOI: 10.19595/j.cnki.1000-6753.tces.220312
电机及其系统 |
基于GAF-CapsNet的电机轴承故障诊断方法
张辉1,2, 戈宝军1, 韩斌2, 赵丽娜1
1.大型电机电气与传热技术国家地方联合研究中心(哈尔滨理工大学) 哈尔滨 150080;
2.齐齐哈尔大学计算机与控制工程学院 齐齐哈尔 161000
Fault Diagnosis Method of Motor Bearing Based on GAF-CapsNet
Zhang Hui1,2, Ge Baojun1, Han Bin2, Zhao Lina1
1. National Engineering Research Center of Large Electric Machines and Heat Transfer Technology Harbin University of Science and Technology Harbin 150080 China;
2. College of Computer and Control Engineering Qiqihar University Qiqihar 161000 China
全文: PDF (7228 KB)   HTML
输出: BibTeX | EndNote (RIS)      
摘要 针对一维机械振动信号在输入卷积神经网络时无法充分提取相对位置关系的问题,提出一种基于格拉姆角场(GAF)和小尺寸卷积的胶囊网络的轴承故障诊断分类方法。利用GAF对采集到的振动信号进行编码,可以很容易地进行角度透视,从而识别出不同时间间隔内的时间相关性并产生相应特征图。胶囊网络对小尺寸图像相对位置比较敏感,特征提取具有优势,同时考虑到VGG网络优秀的特征提取能力,在结合胶囊网络和VGG网络的基础上,加入深度小尺寸卷积层。将GAF编码的振动图像输入到改进的CapsNet网络进行训练,组成GAF-CapsNet模型对轴承故障进行诊断。该模型在凯斯西储大学轴承数据集上进行试验,结果表明,格拉姆角和场(GADF)编码方式相比格拉姆角差场(GASF)编码效果差,效果较好的GADF-CapsNet有99.27 %准确率,较差的GASF-CapsNet也有98.83 %准确率,相较其他编码方式和卷积神经网络,该模型性能表现普遍比其他模型具有更高准确率。
服务
把本文推荐给朋友
加入我的书架
加入引用管理器
E-mail Alert
RSS
作者相关文章
张辉
戈宝军
韩斌
赵丽娜
关键词 轴承故障诊断格拉姆角场胶囊网络    
Abstract:In mechanical motor faults, up to 30 % of the damage is caused by bearing faults, which makes bearing fault diagnosis and maintenance more critical. Traditional intelligent fault diagnosis methods are challenging to achieve good results when dealing with big data because of their poor generalization ability of feature extraction. In recent years, due to the substantial increase in training resources and the rapid development of computing power, deep learning has gradually become a new player in intelligent fault diagnosis.
This paper proposes a new GAF-CapsNet model to solve the problem that the relative position relationship cannot be fully extracted when one-dimensional mechanical vibration signals are input into the convolutional neural network. The Gramian Angular Field (GAF) encoding method converts the original data into images with conspicuous features. The Gramian Angular Summation Fields (GASF) and Gramian Angular Difference Fields (GADF) are used, respectively. Two groups of feature maps are input into the convolution layer of the small convolution kernel for information reading and feature extraction and then into the capsule network for deeper feature extraction and fault identification. Finally, ten capsules of the digital capsule layer correspond to different fault types.
The Gram-angle field encodes the collected vibration signals, which can be quickly perspective to identify the temporal correlations in different time intervals and generate corresponding feature maps. The Capsule network is sensitive to the relative position of small-size images and has advantages in feature extraction. At the same time, considering the excellent feature extraction ability of the VGG network, a deep small-size convolutional layer is added based on the combination of the capsule network and the VGG network. The vibration images encoded by the Gram Angle field were input to the improved CapsNet network for training, and the GAF-CapsNet model was formed to diagnose bearing faults. Among the two methods of GAF coding, GADF coding performs better than GASF coding in a capsule network. Overall, the GAF encoding method retains relatively complete fault characteristics of the original vibration signal. Due to the influence of different sampling sizes on the accuracy, the experiment proves that the 128 sampling size is the best input size for improving the capsule network.
The performance of the GAF-CapsNet model is tested on the rolling bearing data in the bearing fault database of Case Western Reserve University (CWRU). The results show that the GASF coding method has a poor effect compared with the GADF coding method. Gadfly-cabinet with a sound effect has 99.27 % accuracy, and GASF-CAPSNet with a poor effect has 98.83 %. Compared with other coding methods and convolutional neural networks, the performance of this model is generally higher than that of other models. The maximum difference was 1.64 %.
Finally, the proposed model can accurately predict the fault location and severity in the confusion matrix experiment. Compared with one-dimensional convolution and other networks, the proposed model performs strongly in the anti-noise experiment. The model's accuracy can reach more than 65 % when the SNR is -4dB.
Key wordsBearing    fault diagnosis    Gramian angular field    capsule network   
收稿日期: 2022-03-07     
PACS: TM307  
基金资助:国家自然科学基金(51777048)和黑龙江省普通高校基本科研业务专项资金(145209409)资助项目
通讯作者: 戈宝军 男,1960年生,教授,博士生导师,研究方向为电机基础理论与应用。E-mail: gebj@hrbust.edu.cn   
作者简介: 张 辉 男,1982年生,博士研究生,副教授,研究方向为电机智能化与故障诊断。E-mail: zhanghui_zdh@163.com
引用本文:   
张辉, 戈宝军, 韩斌, 赵丽娜. 基于GAF-CapsNet的电机轴承故障诊断方法[J]. 电工技术学报, 2023, 38(10): 2675-2685. Zhang Hui, Ge Baojun, Han Bin, Zhao Lina. Fault Diagnosis Method of Motor Bearing Based on GAF-CapsNet. Transactions of China Electrotechnical Society, 2023, 38(10): 2675-2685.
链接本文:  
https://dgjsxb.ces-transaction.com/CN/10.19595/j.cnki.1000-6753.tces.220312          https://dgjsxb.ces-transaction.com/CN/Y2023/V38/I10/2675