Motor Fault Diagnosis Method Based on Migration Learning and CNN
-
摘要: 针对缺乏数据导致卷积神经网络(Convolutional neural networks,CNN)训练不佳的问题,以三相异步电机故障诊断为研究对象,提出了基于迁移学习和CNN结合的电机故障诊断方法。首先搭建了电机故障诊断实验平台,通过加速度传感器获取CNN模型带标签数据,通过训练获取预训练模型;然后结合迁移学习得到的预训练模型迁移到目标域,并通过对目标域的带标签数据进行训练以优化CNN参数;最终获得可以对目标域数据有着良好分类能力的新模型,从而实现目标域带标签数据稀少情况下的电机故障诊断工作。通过将该方法与传统CNN、变分模态分解(Variational modal decomposition,VMD)-支持向量机(Support vector machine,SVM)、VMD-K近邻(K nearest neighbor,KNN)以及VMD-BP神经网络等识别模型进行对比验证,结果显示本文提出的迁移CNN模型模式识别方法有更好的识别效果。Abstract: Aiming at the problem that the lack of labeled data will lead to poor training of convolutional neural network (CNN), a motor fault diagnosis method based on the combination of migration learning and CNN is proposed for three-phase asynchronous motor fault diagnosis. Firstly, an experimental platform for motor fault diagnosis is built, the label data of input CNN model is obtained by acceleration sensor, and the pre-trained model is obtained through training. Then, the obtained pre-training model is transferred to the target domain with transfer learning, and a small amount of labeled data in the target domain is cleared for training and fine-tuning parameters, and the CNN parameters are optimized by training the labeled data in the target domain. Finally, a new model with good classification ability for the target domain data is obtained, so as to realize the motor fault diagnosis in the case of scarce labeled data in the target domain. By comparing this method with ordinary CNN, variational modal decomposition (VMD)-support vector machine (SVM), VMD-K nearest neighbor (KNN) and VMD-BP neural network recognition models for validation, the results show that the pattern recognition method of migrating CNN model proposed in this paper has better recognition effect.
-
Key words:
- CNN /
- transfer learning /
- three phase asynchronous motor /
- VMD /
- fault diagnosis
-
表 1 迁移学习和传统机器学习的区别
Table 1. Differences between transfer learning and traditional machine learning
区别 迁移学习 传统机器学习 数据分布 不需要同分布 同分布 标签数据 少量标签数据 大量标签数据 建立模型 可以用训练过的模型 重新建模 表 2 CNN模型参数
Table 2. CNN model parameters
网络层 卷积核
大小步长 滤波器
数量输出
大小参数
个数输入层 1024 × 1 64 卷积层1 16 4 64 256 × 64 1088 最大池化层1 2 2 64 128 × 64 卷积层2 3 1 64 128 × 64 12352 最大池化层2 2 2 64 64 × 64 卷积层3 3 2 128 32 × 128 24704 最大池化层3 2 2 128 16 × 128 卷积层4 3 1 256 16 × 256 98560 全局池化层 16 1 256 256 全连接层 768 1 768 197376 Softmax层 1 3 2307 表 3 A→B迁移数据集
Table 3. A→B migration dataset
迁移
任务源域
转速目标域
转速源域样本
个数目标域样本
个数A→B 600 900 1800 1836 B→A 900 600 1800 1836 A→C 600 1200 1800 1836 C→A 1200 600 1800 1836 B→C 900 1200 1800 1836 C→B 1200 900 1800 1836 表 4 不同冻结方法测试对比
Table 4. Comparison of tests using different freezing methods
方法 平均识别率/% 单次模型训练时间/s 不冻结 98.04 1.79 冻结卷积层1 95.82 1.71 冻结卷积层2 93.62 1.68 冻结卷积层3 91.38 1.51 冻结卷积层4 87.43 1.31 冻结全连接层 78.79 1.02 冻结Softmax层 60.17 0 表 5 A→B情况下迁移学习模型指标
Table 5. Transfer learning model indicators in the A→B case
参数
精确率
召回率
F1值
数量
正常
0.9932
0.975
0.984
600
转子断条
0.9781
0.9933
0.9856
600
轴承故障
0.9967
1
0.9983
600表 6 不同方法对比
Table 6. Comparison of different methods
方法 平均总体识别率/% A→B B→A A→C C→A B→C C→B 迁移CNN 97.39 98.43 95.92 98.38 96.71 96.73 传统CNN 90.31 86.4 81.45 86.4 81.45 90.31 VMD-SVM 83.61 89.72 84.27 89.72 84.27 83.61 VMD-KNN 86.11 82.5 78.32 82.5 78.32 86.11 VMD-BP 81.56 77.63 73.70 77.63 73.70 81.56 -
[1] 符羽. 基于卷积神经网络的三相异步电机故障诊断方法研究[D]. 南昌: 华东交通大学, 2021.FU Y. Fault diagnosis of three phase induction motor based on convolution neural network[D]. Nanchang: East China Jiaotong University, 2021. (in Chinese) [2] 郝鹏永. 基于深度信念网络的异步电机典型故障的诊断方法及其实验研究[D]. 秦皇岛: 燕山大学, 2019.HAO P Y. Diagnostic method and experimental study of typical faults of asynchronous motor based on deep belief network[D]. Qinghuangdao: Yanshan University, 2019. (in Chinese) [3] 贺珂珂. 基于深度学习理论的电机故障诊断方法研究[D]. 兰州: 兰州理工大学, 2019.HE K K. Research on motor fault diagnosis method based on deep learning theory[D]. Lanzhou: Lanzhou University of Technology, 2019. (in Chinese) [4] HINTON G E, SALAKHUTDINOV R R . Reducing the dimensionality of data with neural networks[J]. Science,2006 ,313 (5786 ):504 -507 .HINTON G E, SALAKHUTDINOV R R. Reducing the dimensionality of data with neural networks[J]. Science, 2006, 313(5786): 504-507.[5] 仝钰, 庞新宇, 魏子涵. 基于GADF-CNN的滚动轴承故障诊断方法[J]. 振动与冲击, 2021, 40(5): 247-253.TONG Y, PANG X Y, WEI Z H. Fault diagnosis method of rolling bearing based on GADF-CNN[J]. Journal of Vibration and Shock, 2021, 40(5): 247-253. (in Chinese) [6] 丁承君, 冯玉伯, 王曼娜. 基于变分模态分解与深度卷积神经网络的滚动轴承故障诊断[J]. 振动与冲击, 2021, 40(2): 287-296.DING C J, FENG Y B, WANG M N. Rolling bearing fault diagnosis using variational mode decomposition and deep convolutional neural network[J]. Journal of Vibration and Shock, 2021, 40(2): 287-296. (in Chinese) [7] 雷亚国, 贾峰, 周昕, 等. 基于深度学习理论的机械装备大数据健康监测方法[J]. 机械工程学报, 2015, 51(21): 49-56. doi: 10.3901/JME.2015.21.049LEI Y G, JIA F, ZHOU X, et al. A deep learning-based method for machinery health monitoring with big data[J]. Journal of Mechanical Engineering, 2015, 51(21): 49-56. (in Chinese) doi: 10.3901/JME.2015.21.049 [8] 孙文珺, 邵思羽, 严如强. 基于稀疏自动编码深度神经网络的感应电动机故障诊断[J]. 机械工程学报, 2016, 52(9): 65-71. doi: 10.3901/JME.2016.09.065SUN W J, SHAO S Y, YAN R Q. Induction motor fault diagnosis based on deep neural network of sparse auto-encoder[J]. Journal of Mechanical Engineering, 2016, 52(9): 65-71. (in Chinese) doi: 10.3901/JME.2016.09.065 [9] 王丽华, 谢阳阳, 张永宏, 等. 采用深度学习的异步电机故障诊断方法[J]. 西安交通大学学报, 2017, 51(10): 128-134. doi: 10.7652/xjtuxb201710021WANG L H, XIE Y Y, ZHANG Y H, et al. A fault diagnosis method for asynchronous motor using deep learning[J]. Journal of Xi′an Jiaotong University, 2017, 51(10): 128-134. (in Chinese) doi: 10.7652/xjtuxb201710021 [10] 陈祝云. 基于深度迁移学习的机械设备智能诊断方法研究[D]. 广州: 华南理工大学, 2020.CHEN Z Y. Research on intelligent diagnosis of machinery equipment based on deep transfer learning[D]. Guangzhou: South China University of Technology, 2020. (in Chinese) [11] 陈超, 沈飞, 严如强. 改进LSSVM迁移学习方法的轴承故障诊断[J]. 仪器仪表学报, 2017, 38(1): 33-40. doi: 10.3969/j.issn.0254-3087.2017.01.005CHEN C, SHEN F, YAN R Q. Enhanced least squares support vector machine-based transfer learning strategy for bearing fault diagnosis[J]. Chinese Journal of Scientific Instrument, 2017, 38(1): 33-40. (in Chinese) doi: 10.3969/j.issn.0254-3087.2017.01.005 [12] 康守强, 胡明武, 王玉静, 等. 基于特征迁移学习的变工况下滚动轴承故障诊断方法[J]. 中国电机工程学报, 2019, 39(3): 764-772. doi: 10.13334/J.0258-8013.PCSEE.180130KANG S Q, HU M W, WANG Y J, et al. Fault diagnosis method of a rolling bearing under variable working conditions based on feature transfer learning[J]. Proceedings of the CSEE, 2019, 39(3): 764-772. (in Chinese) doi: 10.13334/J.0258-8013.PCSEE.180130 [13] 沈飞, 陈超, 严如强. 奇异值分解与迁移学习在电机故障诊断中的应用[J]. 振动工程学报, 2017, 30(1): 118-126. doi: 10.16385/j.cnki.issn.1004-4523.2017.01.016SHEN F, CHEN C, YAN R Q. Application of SVD and transfer learning strategy on motorfault diagnosis[J]. Journal of Vibration Engineering, 2017, 30(1): 118-126. (in Chinese) doi: 10.16385/j.cnki.issn.1004-4523.2017.01.016 [14] 朱光磊. 三相笼型异步电机故障时的定子电流特征分析与实验研究[D]. 广州: 华南理工大学, 2018.ZHU G L. Study of the current characteristics in three-phase squirrel cage asynchronous motor fault state and experimental verification[D]. Guangzhou: South China University of Technology, 2018. (in Chinese) [15] 高学, 王有旺. 基于CNN和随机弹性形变的相似手写汉字识别[J]. 华南理工大学学报(自然科学版), 2014, 42(1): 72-76.GAO X, WANG Y W. Recognition of similar handwritten Chinese characters based on CNN and random elastic deformation[J]. Journal of South China University of Technology (Natural Science Edition), 2014, 42(1): 72-76. (in Chinese) [16] 于洋, 何明, 刘博, 等. 基于TL-LSTM的轴承故障声发射信号识别研究[J]. 仪器仪表学报, 2019, 40(5): 51-59.YU Y, HE M, LIU B, et al. Research on acoustic emission signal recognition of bearing fault based on TL-LSTM[J]. Chinese Journal of Scientific Instrument, 2019, 40(5): 51-59. (in Chinese) [17] YANG Y H, SHI G Q, SHI X. Fault monitoring and classification of rotating machine based on PCA and KNN[C]//2018 Chinese Control and Decision Conference. Shenyang: IEEE, 2018: 1795-1800.