作者：冯涛, 王杰, 方夏, 刘剑歌, 黄思思

作者单位：四川大学机械工程学院, 四川 成都 610041

关键词：振动马达;故障判别;时频特征图;卷积神经网络

摘要：

为解决人耳听音判别微型振动马达故障困难的局面，提出基于CNN和声音时频特征图的微型马达故障判别方法。通过采集微型振动马达运转过程中的声音信号，进行短时傅里叶变换获得二维时频特征灰度图。将通过经验人员反复听音和相关设备辨别的工件制作成训练集和测试集，通过CNN对训练集中时频特征图进行学习，使网络模型能够具有马达故障判别功能，并在测试集上进行验证。在训练集准确率为99.2%时，测试集准确率为94.1%。为验证模型在实际坏件判别中的可靠性，对6种单一破坏的零件进行分类，平均判别准确率达90%。结果表明：基于CNN和声音时频特征图的微型马达故障判别方法在微型振动马达的故障判别上有可靠的效果，能够运用于工业环境中取代传统的人耳听音判别故障的方法。

Fault diagnosis method for micro-vibration motor based on CNN and time-frequency characteristic map of sound
FENG Tao, WANG Jie, FANG Xia, LIU Jian'ge, HUANG Sisi
School of Mechanical Engineering, Sichuan University, Chengdu 610041, China
Abstract: It is difficult for human ear to distinguish the fault of micro vibration motor. In order to solve the problem, a fault diagnosis method based on convolutional neural network and time-frequency characteristic map of sound has been proposed. The two-dimensional time-frequency characteristic gray scale map was obtained by collecting the sound signal of the operational micro-vibration motor and applying short-time Fourier transform. By learning the time-frequency characteristic map of the training set through CNN, the network model obtained the function of motor fault detection. And the reliability of the network model was verified on the test set. While the training set and test set were made by experienced human and related equipments. The result shows that the accuracy of test set is 94.1% when the accuracy of training set is 99.2%, To verify the reliability of the model on the discrimination of actually bad pieces, the 6 kinds of piece damaged were respectively carried out.The result indicates that the average detection accuracy is up to 90%. This study demonstrates that the fault detection method based on CNN and time-frequency feature map of sound is effective for the fault detection of micro-motor, which can be used in industrial field to replace the traditional human ear-hearing detection method.
Keywords: vibration motor;fault detection;time-frequency characteristic map;convolutional neural network
2019, 45(10):120-127  收稿日期: 2019-01-25;收到修改稿日期: 2019-03-04
基金项目: 四川省科技计划资助(2019YFG0356)；四川省科技厅重点研发项目(2019YFG0359)
作者简介: 冯涛(1994-),男,四川南充市人,硕士研究生,专业方向为智能检测以及深度学习算法应用
参考文献
[1] 宗荣珍. 振动电机偏心块动态特性研究[J]. 煤矿机械, 2010, 31(7):47-49
[2] LI C, FENGLIN L. Application of wavelet analysis in fault detection of cell phone vibration motor[C]//2009 International Asia Conference on Informatics in Control, Automation, and Robotics, 2009.
[3] 丘寿玉. 一种手机振动马达振动量评测方法浅析[J]. 电子质量, 2018(3):15-18
[4] 古莹奎, 潘高平, 朱繁泷, 等. 基于邻域属性重要度与主成分分析的齿轮箱故障特征约简[J]. 中国机械工程, 2016, 27(13):1783-1789
[5] 赵越, 殷鸣, 黄沁元, 等. 基于WPT-ANN的磁瓦内部缺陷音频检测[J]. 中国测试, 2015, 41(6):81-85
[6] 谢罗峰, 徐慧宁, 黄沁元, 等. 应用双树复小波包和NCA-LSSVM检测磁瓦内部缺陷[J]. 浙江大学学报(工学版), 2017, 51(1):184-191
[7] KERBOUA A, METATLA A, KELAIAIA R, et al. Real-time safety monitoring in the induction motor using deep hierarchic long short-term memory[J]. The International Journal of Advanced Manufacturing Technology, 2018, 99(9-12):2245-2255
[8] 张孟伯, 王伦文, 冯彦卿. 基于卷积神经网络的OFDM频谱感知方法[J]. 系统工程与电子技术, 2019, 41(1):178-186
[9] 张振华, 井栋. 基于声音振动信号分析的产品品质测试[J]. 战术导弹技术, 2011(2):115-118
[10] 唐求, 贾杨威, 滕召胜, 等. 基于粒子群优化的声级计A计权设计[J]. 仪器仪表学报, 2015, 36(4):856-862
[11] 刘红星, 左洪福, 姜澄宇, 等. 信号频谱的二维向量及其应用[J]. 中国机械工程, 1999(5):537-539
[12] 熊星, 孔凡让, 张海滨, 等. 基于WPD-HHT的滚动轴承故障诊断[J]. 机械与电子, 2014(1):8-11
[13] 李岳, 韩宾, 鲁云. 基于声卡和LabVIEW的声音信号EMD时频分析系统[J]. 微型机与应用, 2016, 35(7):73-75
[14] 刘炳集. 基于时频图和CNN的直升机自动倾斜器滚动轴承故障诊断方法研究[D]. 南昌:南昌航空大学, 2018.
[15] KRIZHEVSKY A, SUTSKEVER I, HINTON G E. ImageNet classification with deep convolutional neural networks[C]//International Conference on Neural Information Processing Systems. Amsterdam:Springer International Publishing, 2012.