改进条件对抗网络在小样本故障诊断中的研究

谢由生; 张军

doi:10.13433/j.cnki.1003-8728.20220135

改进条件对抗网络在小样本故障诊断中的研究

doi: 10.13433/j.cnki.1003-8728.20220135

谢由生,
张军^,

安徽理工大学人工智能学院，安徽淮南　232001

基金项目: 国家创新方法工作专项（2018IM010500）与安徽省科技重大专项计划项目（16030901012）

详细信息

作者简介:
谢由生（1995−），硕士，研究方向为故障诊断，xie-ys@foxmail.com

通讯作者:
张军，教授，硕士生导师，zhj63@163.com

中图分类号: TG156
计量
- 文章访问数: 106
- HTML全文浏览量: 72
- PDF下载量: 16
- 被引次数: 0
出版历程
- 收稿日期: 2021-09-27
- 刊出日期: 2023-11-30

Research on Improved Conditional Generative Adversarial Network for Small Sample Fault Diagnosis

XIE Yousheng,
ZHANG Jun^,

School of Artificial Intelligence, Anhui University of Science and Technology, Huainan 232001, Anhui, China

摘要

摘要: 在实际智能设备的故障诊断中，往往很难获得大量的故障样本，这对基于机器学习的故障诊断的分类精度造成不可估量的影响。为了提高小样本情况下的故障诊断精度,提出一种基于条件对抗网络的生成模型（Conditional generative adversarial networks-gradient penalty, CGAN-GP）,用于数据增强来获得充足的故障样本。CGAN-GP利用二维卷积，学习预处理后获得的二维故障样本的分布特性，生成与真实样本相似的样本，并使用Wasserstein距离和梯度惩罚(Gradient penalty, GP)策略解决模型训练中的问题，同时将故障样本的标签信息输入模型引导模型生成特定的故障样本，实现一个模型可生成多种故障样本，并且在CWRU轴承数据集上得以验证。研究表明提出的模型可以生成与真实样本特征相似的高质量样本，能够有效提高小样本情况下故障诊断的识别率。
- 故障诊断 /
- 生成对抗网络 /
- 小样本 /
- 梯度惩罚
Abstract: In the fault diagnosis of practical intelligent equipment, it is often difficult to obtain a large number of fault samples, which has an inestimable impact on the classification accuracy of fault diagnosis based on machine learning. In order to improve the accuracy of fault diagnosis in the case of small samples, a generation model (AGAN-GP) based on conditional generative adversarial networks was proposed for data enhancement to obtain sufficient fault samples. AGAN-GP used two-dimensional convolution to learn the distribution characteristics of two-dimensional fault samples obtained after preprocessing, generated samples similar to real samples, used Wasserstein distance and gradient penalty (GP) strategy to solve the problems in model training, and input the label information of fault samples into the model to guide the model to generate specific fault samples. Implementing a model can generate a variety of fault samples and be verified on CWRU bearing data set. The research result shows that the proposed model can generate high-quality samples with similar characteristics to the real samples, and can effectively improve the recognition rate of fault diagnosis in the case of small samples.
- fault diagnosis /
- generative adversarial networks(GAN) /
- small samples /
- gradient penalty

HTML全文

图 1 GAN和CGAN模型结构

Figure 1. GAN and CGAN model structures

下载: 全尺寸图片幻灯片

图 2 CGAN-GP模型训练流程

Figure 2. CGAN-GP model training process

下载: 全尺寸图片幻灯片

图 3 故障诊断流程

Figure 3. Fault diagnosis process

下载: 全尺寸图片幻灯片

图 4 CWRU轴承实验平台^[14]

Figure 4. CWRU bearing experimental platform

下载: 全尺寸图片幻灯片

图 5 真实样本转换后的灰度图

Figure 5. Gray scale image after real sample conversion

下载: 全尺寸图片幻灯片

图 6 CGAN-GP中生成模型的训练损失

Figure 6. Training loss of the model generated in CGAN-GP

下载: 全尺寸图片幻灯片

图 7 真实样本与生成样本对比

Figure 7. Comparison between real samples and generated samples

下载: 全尺寸图片幻灯片

图 8 t-sne可视化生成样本和真实样本

Figure 8. t-sne visualization-generated samples and real samples

下载: 全尺寸图片幻灯片

图 9 故障诊断模型在各训练集的损失及在验证集上的准确率

Figure 9. Loss of the fault diagnosis model in each training set and its accuracy on accuracy validation set

下载: 全尺寸图片幻灯片

图 10 各故障诊断模型在测试集上的准确率

Figure 10. Accuracy of each fault diagnosis model on the test set

下载: 全尺寸图片幻灯片

图 11 各故障诊断模型在测试集上诊断结果的混淆矩阵

Figure 11. Confusion matrix of diagnostic results for each fault diagnosis model on the test set

下载: 全尺寸图片幻灯片

表 1 CGAN-GP具体参数

Table 1. Specific parameters of CGAN-GP

层数	生成模型	判别模型
1	全连接层	Conv4_32
2	BatchNorm	LeakyReLU层
3	上采样层	Conv4_64
4	Conv3_128	LeakyReLU层
5	BatchNorm	Conv4_128
6	ReLU层	LeakyReLU层
7	上采样层	Conv4_256
8	Conv3_64	LeakyReLU层
9	BatchNorm	Conv4_1
10	ReLU层
11	上采样层
12	Conv3_64
13	BatchNorm
14	ReLU层
15	上采样层
16	Conv3_1
17	Tanh层

下载: 导出CSV

表 2 故障类别及对应标签

Table 2. Fault categories and corresponding labels

标签	故障位置	故障尺寸/mm
0（b007）	滚动体	0.1800
1（b014）	滚动体	0.3556
2（b021）	滚动体	0.7112
3（ir007）	内圈	0.1800
4（ir014）	内圈	0.3556
5（ir021）	内圈	0.7112
6（normal）	正常	−
7（or007）	外圈	0.1800
8（or014）	外圈	0.3556
9（or021）	外圈	0.7112

下载: 导出CSV

表 3 不同方法的生成样本对比

Table 3. Comparison of generated samples using different methods

标签	CGAN-GP		CGAN
标签	ED	CD	ED	CD
0（b007）	0.537087	0.029969	0.551945	0.035627
2（b021）	0.538866	0.037499	0.560798	0.039045
4（ir014）	0.528053	0.036538	0.608441	0.048488
6（normal）	0.862965	0.101061	0.765576	0.078399
8（or014）	0.562402	0.040803	0.591200	0.046606

下载: 导出CSV

表 4 故障诊断模型实验分组

Table 4. Experimental grouping of fault diagnosis model

分组	训练集		模型名称
分组	真实样本	生成样本	模型名称
A	0	100	g100
B	0	400	g400
C	0	600	g600
D	0	800	g800
E	50	0	r50
F	50	100	r50_g100
G	50	400	r50_g400
H	50	600	r50_g600
I	50	800	r50_g800

下载: 导出CSV

参考文献(15)

[1]	CHEN X F, WANG S B, QIAO B J, et al. Basic research on machinery fault diagnostics: past, present, and future trends[J]. Frontiers of Mechanical Engineering, 2018, 13(2): 264-291. doi: 10.1007/s11465-018-0472-3
[2]	叶壮, 余建波. 基于多通道加权卷积神经网络的齿轮箱振动信号特征提取[J]. 机械工程学报, 2021, 57(1): 110-120. doi: 10.3901/JME.2021.01.110 YE Z, YU J B. Feature extraction of gearbox vibration signals based on multi-channels weighted convolutional neural network[J]. Journal of Mechanical Engineering, 2021, 57(1): 110-120. (in Chinese) doi: 10.3901/JME.2021.01.110
[3]	齐咏生, 樊佶, 李永亭, 等. 一种改进的解卷积算法及其在滚动轴承复合故障诊断中的应用[J]. 振动与冲击, 2020, 39(21): 140-150. doi: 10.13465/j.cnki.jvs.2020.21.019 QI Y S, FAN J, LI Y T, et al. An improved deconvolution algorithm and its application in compound fault diagnosis of rolling bearing[J]. Journal of Vibration and Shock, 2020, 39(21): 140-150. (in Chinese) doi: 10.13465/j.cnki.jvs.2020.21.019
[4]	WEN L, LI X Y, GAO L, et al. A new convolutional neural network-based data-driven fault diagnosis method[J]. IEEE Transactions on Industrial Electronics, 2018, 65(7): 5990-5998. doi: 10.1109/TIE.2017.2774777
[5]	SHAO H D, JIANG H K, ZHANG H Z, et al. Rolling bearing fault feature learning using improved convolutional deep belief network with compressed sensing[J]. Mechanical Systems and Signal Processing, 2018, 100: 743-765. doi: 10.1016/j.ymssp.2017.08.002
[6]	ZHANG K, TANG B P, DENG L, et al. A hybrid attention improved ResNet based fault diagnosis method of wind turbines gearbox[J]. Measurement, 2021, 179: 109491. doi: 10.1016/j.measurement.2021.109491
[7]	GOODFELLOW I, POUGET-ABADIE J, MIRZA M, et al. Generative adversarial nets[C]//Proceedings of the 27th International Conference on Neural Information Processing Systems. Montreal: MIT Press, 2014: 2672-2680.
[8]	LUO J, HUANG J Y, LI H M. A case study of conditional deep convolutional generative adversarial networks in machine fault diagnosis[J]. Journal of Intelligent Manufacturing, 2021, 32(2): 407-425. doi: 10.1007/s10845-020-01579-w
[9]	SHAO S Y, WANG P, YAN R Q. Generative adversarial networks for data augmentation in machine fault diagnosis[J]. Computers in Industry, 2019, 106: 85-93. doi: 10.1016/j.compind.2019.01.001
[10]	MIRZA M, OSINDERO S. Conditional generative adversarial nets[J]. Computing Research Repository, 2014, abs/1411.1784.
[11]	ARJOVSKY M, CHINTALA S, BOTTOU L. Wasserstein generative adversarial networks[C]//Proceedings of the 34th International Conference on Machine Learning. Sydney: PMLR, 2017: 214-223.
[12]	GULRAJANI I, AHMED F, ARJOVSKY M, et al. Improved training of wasserstein GANs[C]//Proceedings of the 31st International Conference on Neural Information Processing Systems. Long Beach: Curran Associates Inc. , 2017: 5769-5779.
[13]	KINGMA D P, BA J. ADAM: a method for stochastic optimization[C]//Proceedings of the 3rd International Conference on Learning Representations. San Diego: ICLR, 2014.
[14]	SMITH W A, RANDALL R B. Rolling element bearing diagnostics using the Case Western Reserve University data: a benchmark study[J]. Mechanical Systems and Signal Processing, 2015, 64-65: 100-131. doi: 10.1016/j.ymssp.2015.04.021
[15]	VAN DER MAATEN L, HINTON G. Visualizing data using t-SNE[J]. Journal of Machine Learning Research, 2008, 9(86): 2579-2605.