动态增殖流形学习算法在机械故障诊断中的应用  

A dynamic incremental manifold learning algorithm and its application in fault diagnosis of machineries

作　　者：宋涛[1] 汤宝平[1] 邓蕾[1] 

机构地区：[1]重庆大学机械传动国家重点实验室,重庆400044

出　　处：《振动与冲击》2014年第23期15-19,29,共6页Journal of Vibration and Shock

基　　金：国家自然科学基金(51275546);高校博士点专项科研基金(20130191130001)

摘　　要：针对现有的批量式流形学习算法无法利用已学习的流形结构实现新增样本的快速约简的缺点,提出增殖正交邻域保持嵌入(Incremental Orthogonal Neighborhood Preserving Embedding,IONPE)流形学习算法。该算法在正交邻域保持嵌入算法基础上利用分块处理思想实现新增样本子集的动态约简。从原始样本中选取部分重叠点合并至新增样本,对重叠点和新增样本子集不依赖原始样本使用正交邻域保持嵌入(ONPE)进行独立约简获取低维嵌入坐标子集,并基于重叠点坐标差值最小化原则,将新增样本低维嵌入坐标通过旋转平移缩放整合到原样本子集中。齿轮箱故障诊断案例证实了IONPE算法具有良好的增量学习能力,在继承ONPE优良聚类特性的同时有效提高了新增样本约简效率。The current batch manifold learning algorithms can＇t achieve rapid dimension reduction of additional samples with learned manifold structures. Here,the incremental orthogonal neighborhood preserving embedding（ IONPE）manifold learning algorithm was proposed. With it,dynamic incremental learning for additional samples was realized with a block processing idea based on orthogonal neighborhood preserving embedding. Firstly,some overlapping points were selected from the original samples and added to the additional samples. Secondly, the subset of low-dimensional embedding coordinates of additional samples was obtained with ONPE independing on the original samples. Finally,based on the principle of minimizing the differences of the overlapping point coordinates,the low-dimensional embedding coordinates of the additional samples were integrated into the original samples with rotating, shifting and scaling transformations. The fault diagnosis case of a gearbox confirmed that the IONPE algorithm has a good incremental learning ability,it improves the processing efficiency of the additional samples while inheriting the superior clustering performance of ONPE.

关 键 词：增殖流形学习 正交邻域保持嵌入 动态约简 分块处理 故障诊断 

分 类 号：TH165.3[机械工程—机械制造及自动化]