数据特征选择与分类在机械故障诊断中的应用  

Application of data feature selection and classification in mechanical fault diagnosis

作　　者：李帅位 张栋良 黄昕宇 陈璞 LI Shuaiwei;ZHANG Dongliang;HUANG Xinyu;CHEN Pu(School of Automation Engineering,Shanghai University of Electric Power,Shanghai 200090,China)

机构地区：[1]上海电力学院自动化工程学院

出　　处：《振动与冲击》2020年第2期218-222,共5页Journal of Vibration and Shock

基　　金：国家自然科学基金(61503237);上海市自然科学基金(15ZR1418300);上海市科研计划项目(18020500900)

摘　　要：针对机械故障数据的高维性和不平衡性,提出基于格拉斯曼流形的多聚类特征选择和迭代近邻过采样的故障分类方法。对采集到的振动信号,提取时域和频域相关特征,利用多聚类特征选择将高维数据以局部流形结构映射到低维特征集合。无标签样本借助迭代近邻过采样以恢复最大平衡性为目标进行样本分类,并对剩余无标签样本进行模糊分类。选取滚动轴承正常、外圈、内圈以及滚动体的故障数据,并与支持向量机、基于图的半监督学习算法进行对比。结果表明,提出的方法能有效识别出少数类故障,并在整体上有显著的分类效果。Aiming at the high dimensionality and unbalance property of mechanical failure data,a fault classification method was proposed based on the Grassmann multi-clustering feature selection and iteration nearest neighbor oversampling.Time domain and frequency domain correlation characteristics were extracted from the collected vibration signals,and the high-dimensional data were mapped to low-dimensional feature sets by the multi-clustering feature selection.The unlabeled samples were classified by the iteration nearest neighbor oversampling to restore the data maximum balance,then the remaining unlabeled samples were classified by fuzzy classification.The data of a rolling bearing in normal,state as well as in the states of outer ring,inner ring and rolling elements failure respectively were selected to do an illustrative fault diagnosis.Comparing with the method of support vector machine and semi-supervised learning algorithm based on graph,the results show that the method proposed can effectively identify minority class of faults and has a significant classification effect on the whole diagnosis process.

关 键 词：格拉斯曼流形 多聚类特征选择 迭代近邻过采样 模糊分类 半监督学习 

分 类 号：TH165.3[机械工程—机械制造及自动化] TH133.3