Unbalanced convolution network diagnosis of bearing faults under imbalanced samples condition
-
摘要: 为了提高样本不均衡条件下轴承故障诊断精度,提出了基于VAE-SNN的样本增广方法和基于非平衡损失网络的故障诊断方法。首先,使用变分自编码器用于数据生成,并依据孪生神经网络对生成数据的类别进行判定,实现了基于变分自编码器和孪生神经网络的样本增广;其次,分析了卷积神经网络无差别对待样本的缺点,针对不均衡样本的特殊性,提出了非平衡损失函数卷积网络,该网络能够自动关注数量少、难分的样本训练。经实验验证,生成对抗网络增广的样本相似度为0.847,孪生神经网络增广的样本相似度比对抗网络提高了6.61%,说明孪生神经网络的样本增广效果更好;在相同诊断方法前提下,样本增广后比增广前的准确率提高了9.42%,说明样本增广有利于提高轴承的故障诊断准确率;非均衡损失网络比卷积神经网络的诊断精度提高了7.17%,比自适应深度学习提高了4.12%,验证了非均衡损失网络的高准确率和优越性。Abstract: In order to improve the accuracy of bearing fault diagnosis under the condition of unbalanced samples, a sample expansion method based on VAE-SNN and a fault diagnosis method based on unbalanced loss network are proposed. Firstly, the variational self-coder is used to generate data, and the category of generated data is determined according to the twin neural network. The sample expansion based on the variational self-coder and the twin neural network is realized; Secondly, the disadvantages of convolutional neural network are analyzed. Aiming at the particularity of unbalanced samples, an unbalance loss function convolutional network is proposed, which can automatically focus on the training of small and difficult samples. The experimental results show that the similarity of the expanded samples of the generated countermeasure network is 0.847, and the similarity of the expanded samples of the twin neural network is 6.61% higher than that of the countermeasure network, indicating that the sample expansion effect of the twin neural network is better; On the premise of the same diagnosis method, the accuracy rate of the expanded sample is increased by 9.42% compared with that before the expansion, which indicates that the expanded sample is useful to improve the accuracy rate of bearing fault diagnosis; The diagnosis accuracy of unbalanced loss network is 7.17% higher than that of convolutional neural network and 4.12% higher than that of adaptive deep learning, which verifies the high accuracy and superiority of unbalanced loss network.
-
表 1 轴承故障特征频率
故障位置 特征频率 内圈故障 ${f_{{\rm{IR}}} } = \dfrac{Z}{2}\left( {1 + \dfrac{d}{D}\cos \alpha } \right){f_r}$ 滚动体故障 ${f_{\rm{B}}} = \dfrac{D}{ {2d} }\left( {1 - { {\left( {\dfrac{d}{D} } \right)}^2}\cos \alpha } \right){f_r}$ 外圈故障 ${f_{{\rm{OR}}} } = \dfrac{Z}{2}\left( {1 - \dfrac{d}{D}\cos \alpha } \right){f_r}$ 表 2 非平衡损失网络结构参数
网络名称 网络结构 输出维度 输入层 1 000×1 1 000×1 卷积层1 卷积层参数(1,16,5,4) 16×249 池化层1 池化层参数(3,2) 16×124 卷积层2 卷积层参数(16,16,5,2) 16×62 池化层2 池化层参数(3,3) 16×20 全连接层1 神经元数(320,64) 64 全连接层2 神经元数(64,5) 5 表 3 样本增广相似度
样本 对抗网络增广相似度 孪生网络增广相似度 滚动体0.014"故障 0.887 0.916 外圈0.007"故障 0.835 0.896 内圈0.014"故障 0.851 0.926 外圈0.014"故障 0.815 0.899 整体样本 0.847 0.903 表 4 不同方案的诊断准确率
方法 诊断准确率/(%) 非均衡样本+非平衡损失网络 89.48 增广样本+卷积网络 91.73 增广样本+自适应深度学习 94.78 增广样本+非平衡损失网络 98.90 -
[1] 刘前进, 高丙朋, 宋振军, 等. 基于参数优化变分模态分解的变速工况下轴承故障诊断[J]. 轴承, 2022(8): 71-78. [2] Alabsi M, Liao Y, Nabulsi A A. Bearing fault diagnosis using deep learning techniques coupled with handcrafted feature extraction: a comparative study[J]. Journal of Vibration and Control, 2021, 27(3-4): 404-414. doi: 10.1177/1077546320929141 [3] 朵慕社, 纪国宜. 基于Hilbert-Huang变换和残差神经网络的轴承故障诊断[J]. 制造业自动化, 2022, 44(6): 20-23. [4] Soother D K, Kalwar I H, Hussain T, et al. A novel method based on UNET for bearing fault diagnosis[J]. Computers, Materials and Continua, 2021, 69(1): 393-408. doi: 10.32604/cmc.2021.014941 [5] 雷高伟, 张清华, 苏乃权, 等. 基于深度学习与电子听诊器的轴承故障诊断[J]. 机床与液压, 2022, 50(9): 210-214. [6] 曹梦婷, 谷玉海, 王红军, 等. 基于GRU与迁移学习的滚动轴承故障诊断[J]. 现代制造工程, 2022(1): 143-147. [7] Neupane D, Seok J. Bearing fault detection and diagnosis using case western reserve university dataset with deep learning approaches: A review[J]. IEEE Access, 2020, 8: 93155-93178. [8] 金江涛, 许子非, 李春, 等. 基于深度学习与混沌特征融合的滚动轴承故障诊断[J]. 控制理论与应用, 2022, 39(1): 109-116. [9] 王亚辉, 刘德平, 王宇. 基于GSA-VMD和自适应CNN的滚动轴承故障诊断[J]. 组合机床与自动化加工技术, 2022(7): 85-89. [10] Li B, Tong R, Kang J, et al. Bearing fault diagnosis using synthetic quantitative index-based adaptive underdamped stochastic resonance[J]. Mathematical Problems in Engineering, 2021, 2021(19): 1-10. [11] 张中洋. 少样本条件下的滚动轴承故障诊断方法研究[D]. 南京: 南京信息工程大学, 2022. [12] Harefa E, Zhou W. Performing sequential forward selection and variational autoencoder techniques in soil classification based on laser-induced breakdown spectroscopy[J]. Analytical Methods:Advancing Methods and Applications, 2021, 13(41): 4926-4933. [13] 许泽伟, 彭珍瑞, 张亚峰, 等. 基于多项式混沌展开和KL散度的随机有限元模型修正[J]. 机械强度, 2021, 43(6): 1297-1302. [14] Kingma D P, Salimans T, Welling M. Variational dropout and the local reparameterization trick [C]. Proceedings of the 28th International Conference on Neural Information Processing Systems.US:MIT Press,2015. [15] 刘岱, 常东润, 孙习习, 等. 基于卷积孪生神经网络的滚动轴承故障定位方法[J]. 机电工程, 2022, 39(03): 309-316. doi: 10.3969/j.issn.1001-4551.2022.03.004 [16] Zhang X, Nie J, Zong L, et al. One shot learning with margin[C]//Pacific-Asia Conference on Knowledge Discovery and Data Mining. Macau: Springer, Cham, 2019: 305-317. [17] 李中, 卢春华, 王星, 等. 考虑滚动轴承故障位置与损伤程度的双分支卷积神经网络故障诊断方法[J]. 科学技术与工程, 2022, 22(4): 1441-1448. [18] 王琦, 邓林峰, 赵荣珍. 基于改进一维卷积神经网络的滚动轴承故障识别[J]. 振动与冲击, 2022, 41(3): 216-223. [19] 张中洋. 少样本条件下的滚动轴承故障诊断方法研究[D]. 南京: 南京信息工程大学, 2022 [20] 贾峰, 李世豪, 沈建军, 等. 采用深度迁移学习与自适应加权的滚动轴承故障诊断[J]. 西安交通大学学报, 2022, 56(8): 1-10.