基于卷积神经网络图像分类的轴承故障模式识别

针对传统的基于数据驱动的机械故障模式识别方法中需要人工构造算法提取特征以及人工构造特征提取算法繁琐的问题,结合卷积神经网络(CNN)在图像特征自动提取与图像分类识别中的广泛应用,提出了一种基于CNN图像分类的轴承故障模式识别方法。首先,利用集合经验模态分解(EEMD)方法对轴承振动信号进行自适应分解并用相关系数对得到的本征模函数分量进行筛选。其次,对筛选得到的本征模函数分量进行伪魏格纳-威利时频分析(PWVD)计算得到信号的时频分布图,并对时频图进行预处理。最后,将轴承15种不同工况预处理后的时频图利用CNN进行特征提取与分类识别。将该方法与同类方法进行了对比,分类正确率提高了4.26%。

Bearing fault pattern recognition based on image classification with CNN

Conventional machinery fault pattern recognition methods based on data-driven need to construct  algorithms manually to extract characteristics. Manually constructed  feature  extraction  algorithms  are  complicated. In order  to  solve the problems , a  bearing  fault  pattern  recognition  method  based  on  image  classification with CNN  was  proposed, with  the  consideration  of  the  wide  application  of  CNN  in  image  feature  automatic extraction and classification. Firstly, with the EEMD method, bearing  vibration  signals  were  decomposed  and  a  finite  number  of  intrinsic  mode  functions(IMF)  were  selected  according  to  the  criterion  of  correlation  coefficient. Secondly, with the PWVD method, time-frequency  diagrams  were  obtained  from  selected  IMFs  and  preprocessing  was  carried  on  time-frequency  diagrams. Lastly, fifteen kinds of preprocessed bearing time-frequency diagrams served as import of CNN for feature extraction and classification. Comparison  was  made  between  this  methodology  and  another  similar  method, and the classification accuracy improved by 4.26%.