半监督TS网络模型在齿轮故障诊断中的应用

陈保家; 阮宇豪; 陈法法; 肖文荣; 李公法; 陶波

doi:10.13433/j.cnki.1003-8728.20230012

半监督TS网络模型在齿轮故障诊断中的应用

doi: 10.13433/j.cnki.1003-8728.20230012

陈保家^{1, 2,},
阮宇豪¹,
陈法法^{1, 2, ,},
肖文荣¹,
李公法³,
陶波³

1.
三峡大学水电机械设备设计与维护湖北省重点实验室，湖北宜昌 443002
2.
重庆大学机械传动国家重点实验室，重庆 400030
3.
武汉科技大学冶金装备及其控制教育部重点实验室，武汉 430081

基金项目:

国家自然科学基金项目 51975324

机械传动国家重点实验室开放基金项目 SKLMT-MSKFKT-202020

水电机械设备设计与维护湖北省重点实验室(三峡大学)开放基金项目 2020KJX02

水电机械设备设计与维护湖北省重点实验室(三峡大学)开放基金项目 2021KJX02

水电机械设备设计与维护湖北省重点实验室(三峡大学)开放基金项目 2021KJX13

武汉科技大学冶金装备及其控制教育部重点实验室开放基金项目 MECOMF2021B04

详细信息

作者简介:
陈保家，教授，博士生导师，博士，cbjia@163.com

通讯作者:
陈法法，教授，博士生导师，博士，chenfafa2005@126.com

中图分类号: HT17
计量
- 文章访问数: 7
- HTML全文浏览量: 5
- PDF下载量: 0
- 被引次数: 0
出版历程
- 收稿日期: 2022-03-18
- 刊出日期: 2024-07-25

Application of Semi-supervised Teacher-Student Network Model in Gear Fault Diagnosis

CHEN Baojia^{1, 2
,},
RUAN Yuhao¹,
CHEN Fafa^{1, 2
, ,},
XIAO Wenrong¹,
LI Gongfa³,
TAO Bo³

1.
Hubei Key Laboratory of Hydroelectric Machinery Design and Maintenance, College of Mechanical and Power Engineering, China Three Gorges University, Yichang 443002, Hubei, China
2.
State Key Laboratory of Mechanical Transmission, Chongqing University, Chongqing 400030, China
3.
Key Laboratory of Metallurgical Equipment and Control, Ministry of Education, Wuhan University of Science and Technology, Wuhan 430081, China

摘要

摘要: 为解决在工业大数据条件下，有标签样本少导致机械故障诊断准确率低的问题，提出了一种半监督神经网络模型。该方法采用协同训练的方式，从时域和频域两个维度训练教师网络(T)，将无标签数据转化为高质量的伪标签数据。再利用转化后的伪标签数据训练学生网络(S)，通过对数据进行评判和计分，避免网络过拟合。最后通过得分函数，对伪标签数据进行阶梯筛选成为有标签数据。齿轮故障诊断结果表明: TS网络在仅有少量有标签数据的情况下，故障分类准确率达90.31%，与其他半监督方法相比，准确率高出15%~20%。在信噪比(SNR)为5、0、-5的条件下，模型可以达到86.81%、78.00%、52.78%的诊断准确率。
- 齿轮故障诊断 /
- 伪标签学习 /
- 抗噪性 /
- 协同训练
Abstract: In practical industrial applications, the fault diagnosis accuracy of mechanical equipment is low due to the small amount of labeled data. A TS neural network model based on semi-supervised learning strategy is proposed to solve this problem. Firstly, the teacher network (T) is trained by co-training from both the time domain and frequency domain, then it is used to convert unlabeled data into high-quality pseudo-labeled data. Secondly, the pseudo-labeled data is used to train the student network (S) and avoid network overfitting by judging and scoring the data. Finally, the pseudo-label data is filtered into the labeled data through the scoring function. The results of gear fault diagnosis show that the TS network has a fault classification accuracy of 90.31%, which is 15%-20% higher than other semi-supervised methods. Under the condition of SNR of 5, 0 and -5, the model can achieve diagnostic accuracy of 86.81%, 78.00% and 52.78%.
- gear fault diagnosis /
- pseudo-label learning /
- anti-noise /
- co-training

HTML全文

图 1 网络关系

Figure 1. Network relationships

下载: 全尺寸图片幻灯片

图 2 计分模块

Figure 2. Scoring module

下载: 全尺寸图片幻灯片

图 3 筛选模块

Figure 3. Filtering module

下载: 全尺寸图片幻灯片

图 4 模型训练流程图

Figure 4. Model training process

下载: 全尺寸图片幻灯片

图 5 实验台展示图

Figure 5. Experimental bench display diagram

下载: 全尺寸图片幻灯片

图 6 计分模块超参数

Figure 6. Scoring module hyperparameters

下载: 全尺寸图片幻灯片

图 7 筛选模块超参数

Figure 7. Filter module hyperparameters

下载: 全尺寸图片幻灯片

图 8 半监督方法准确率对比

Figure 8. Comparison of accuracy of semi supervised methods

下载: 全尺寸图片幻灯片

图 9 模型可视化

Figure 9. Model visualization

下载: 全尺寸图片幻灯片

表 1 网络超参数设置

Table 1. Network hyperparameter settings

网络	T₁	T₂	S
输入层	(1 024, 1)	(512, 1)	(1 024, 1)
卷积层	(16, 8)	(16, 8)	(16, 8)
池化层	4	4	4
卷积层	(64, 4)	(64, 4)	(64, 4)
池化层	2	2	2
全连接层	64	64	128
分类层	18	18	18

下载: 导出CSV

表 2 数据集划分

Table 2. Dataset division

载荷	样本数
载荷	有标签	无标签	测试集
空载	30	300	240
轻载	30	300	240
中载	30	300	240

下载: 导出CSV

表 3 不同样本数准确率比较

Table 3. Comparison of accuracy for different sample numbers

样本数	时域	频域	协同训练
5	10.8%	9.2%	16.3%
10	12.1%	10.4%	22.8%
15	23.9%	12.3%	40.7%
20	34.7%	17.8%	55.6%
25	53.1%	33.8%	79.1%

下载: 导出CSV

表 4 计分模块超参数表

Table 4. Scoring module hyperparameters

超参数	搜索范围
准确率阈值A	[0.6, 0.7, 0.8, 0.9]
分类概率阈值P	[0.5, 0.6, 0.7, 0.8, 0.9]

下载: 导出CSV

表 5 筛选模块超参数表

Table 5. Filter module hyperparameters

超参数	阈值范围
备选数据阈值S₀	[5, 6, 7, 8, 9]
有标签数据阈值S₁	[15, 18, 21, 24]

下载: 导出CSV

表 6 TS网络与有监督学习方法准确率对比

Table 6. Comparison of accuracy between TS network and supervised learning method

样本量	TS网络	CNN	ResNet
90	90.31%	69.83%	81.14%
180	97.6%	88.06%	91.44%
270	97.5%	95.53%	94.36%
360	98.61%	97.92%	94.54%

下载: 导出CSV

表 7 半监督方法准确率表

Table 7. Accuracy of semi-supervised methods

信噪比/dB	TS网络	Noise-student	Self-train	FDD
-5	52.78%	49.33%	38.38%	37.61%
0	78.00%	68.81%	58.60%	60.14%
5	86.81%	76.94%	71.90%	70.84%
无	90.31%	74.81%	72.69%	71.32%

下载: 导出CSV

参考文献(20)

[1]	李益兵, 黄定洪, 马建波, 等. 基于深度置信网络与信息融合的齿轮故障诊断方法[J]. 振动与冲击, 2021, 40(8): 62-69. https://www.cnki.com.cn/Article/CJFDTOTAL-ZDCJ202108008.htm LI Y B, HUANG D H, MA J B, et al. A gear fault diagnosis method based on deep belief network and information fusion[J]. Journal of Vibration and Shock, 2021, 40(8): 62-69. (in Chinese) https://www.cnki.com.cn/Article/CJFDTOTAL-ZDCJ202108008.htm
[2]	吴春志, 吴守军, 冯辅周, 等. 一种具有强抗噪性的深度学习故障诊断模型[J]. 西安交通大学学报, 2021, 55(4): 61-68. https://www.cnki.com.cn/Article/CJFDTOTAL-XAJT202104007.htm WU C Z, WU S J, FENG F Z, et al. A deep learning fault diagnosis model with strong anti-noise ability[J]. Journal of Xi'an Jiaotong University, 2021, 55(4): 61-68. (in Chinese) https://www.cnki.com.cn/Article/CJFDTOTAL-XAJT202104007.htm
[3]	AI Y T, GUAN J Y, FEI C W, et al. Fusion information entropy method of rolling bearing fault diagnosis based on n-dimensional characteristic parameter distance[J]. Mechanical Systems and Signal Processing, 2017, 88: 123-136. doi: 10.1016/j.ymssp.2016.11.019
[4]	LI X Y, LI J L, ZHAO C Y, et al. Gear pitting fault diagnosis with mixed operating conditions based on adaptive 1D separable convolution with residual connection[J]. Mechanical Systems and Signal Processing, 2020, 142: 106740. doi: 10.1016/j.ymssp.2020.106740
[5]	ZHAO R, YAN R Q, CHEN Z H, et al. Deep learning and its applications to machine health monitoring[J]. Mechanical Systems and Signal Processing, 2019, 115: 213-237. doi: 10.1016/j.ymssp.2018.05.050
[6]	陈仁祥, 黄鑫, 杨黎霞, 余伟等. 基于卷积神经网络和离散小波变换的滚动轴承故障诊断[J]. 振动工程学报, 2018, 31(5): 883-891. CHEN R X, HUANG X, YANG L X, et al. Rolling bearing fault identification based on convolution neural network and discrete wavelet transform[J]. Journal of Vibration Engineering, 2018, 31(5): 883-891. (in Chinese)
[7]	陈保家, 陈学力, 沈保明, 公等. CNN-LSTM深度神经网络在滚动轴承故障诊断中的应用[J]. 西安交通大学学报, 2021, 55(6): 28-36. CHEN B J, CHEN X L, SHEN B M, et al. An application of convolution neural network and long short-term memory in rolling bearing fault diagnosis[J]. Journal of Xi'an Jiaotong University, 2021, 55(6): 28-36. (in Chinese)
[8]	TANHA J, SAMADI N, ABDI Y, et al. CPSSDS: Conformal prediction for semi-supervised classification on data streams[J]. Information Sciences, 2022, 584: 212-234. doi: 10.1016/j.ins.2021.10.068
[9]	ADAMS N. Semi-supervised learning[J]. Journal of the Royal Statistical Society: Series A (Statistics in Society), 2009, 172(2): 530.
[10]	刘雅芬, 郑艺峰, 江铃燚, 等. 深度半监督学习中伪标签方法综述[J]. 计算机科学与探索, 2022, 16(6): 1279-1290. https://www.cnki.com.cn/Article/CJFDTOTAL-KXTS202206004.htm LIU Y F, ZHENG Y F, JIANG L Y, et al. Survey on pseudo-labeling methods in deep semi-supervised learning[J]. Journal of Frontiers of Computer Science & Technology, 2022, 16(6): 1279-1290. (in Chinese) https://www.cnki.com.cn/Article/CJFDTOTAL-KXTS202206004.htm
[11]	XIE Q Z, LUONG M T, HOVY E, et al. Self-training with noisy student improves ImageNet classification[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle, WA, USA: IEEE, 2020: 10687-10698.
[12]	ZHANG L, YANG L, MA T W, et al. A self-training semi-supervised machine learning method for predictive mapping of soil classes with limited sample data[J]. Geoderma, 2021, 384: 114809. doi: 10.1016/j.geoderma.2020.114809
[13]	FAN C, LIU X Y, XUE P, et al. Statistical characteriza-tion of semi-supervised neural networks for fault detection and diagnosis of air handling units[J]. Energy and Buildings, 2021, 234: 110733. doi: 10.1016/j.enbuild.2021.110733
[14]	PHAM H, DAI Z H, XIE Q Z, et al. Meta pseudo labels[J/OL]. arXiv: 2003.10580, (2021-03-01). https://arxiv.org/abs/2003.10580.
[15]	KRIZHEVSKY A, SUTSKEVER I, HINTON G E. ImageNet classification with deep convolutional neural networks[C]//Proceedings of the 25th International Conference on Neural Information Processing Systems. Lake Tahoe Nevada: Curran Associates Inc., 2012: 1097-1105.
[16]	LECUN Y, BENGIO Y, HINTON G. Deep learning[J]. Nature, 2015, 521(7553): 436-444. doi: 10.1038/nature14539
[17]	GU J X, WANG Z H, KUEN J, et al. Recent advances in convolutional neural networks[J]. Pattern Recognition, 2018, 77: 354-377. doi: 10.1016/j.patcog.2017.10.013
[18]	朱威, 屈景怡, 吴仁彪. 结合批归一化的直通卷积神经网络图像分类算法[J]. 计算机辅助设计与图形学学报, 2017, 29(9): 1650-1657. https://www.cnki.com.cn/Article/CJFDTOTAL-JSJF201709008.htm ZHU W, QU J Y, WU R B. Straight convolutional neural networks algorithm based on batch normalization for image classification[J]. Journal of Computer-Aided Design & Computer Graphics, 2017, 29(9): 1650-1657. (in Chinese) https://www.cnki.com.cn/Article/CJFDTOTAL-JSJF201709008.htm
[19]	CHENG C, LI C P, HAN Y F, et al. A semi-supervised deep learning image caption model based on Pseudo Label and N-gram[J]. International Journal of Approximate Reasoning, 2021, 131: 93-107.
[20]	ASGHAR S, CHOI J, YOON D, et al. Spatial pseudo-labeling for semi-supervised facies classification[J]. Journal of Petroleum Science and Engineering, 2020, 195: 107834.