摘要
针对传统流形学习方法难以处理大批量设备运行数据的特点,提出了一种采用增量式流形学习方法的机械设备状态监测方法.该方法首先利用局部切空间排列算法对训练样本集进行非线性维数约简,得到初始的低维流形结构,然后通过增量式学习机制对新增的时序样本点进行动态聚类.通过对压缩机喘振试验数据及滚动轴承故障数据的分析表明,该方法的计算复杂度低,可以有效地识别出隐藏在高维特征集中的非线性故障特征,因此具有良好的工程应用前景.
Most nonlinear manifold learning methods can not be efficiently operated in a ＇batch＇ mode when data are collected sequentially. An incremental learning scheme is proposed for condition monitoring of mechanical equipments. A nonlinear dimensionality reduction algorithm called local tangent space alignment （LTSA） is first utilized to train the low-dimensional manifold struc- ture from the original feature space. To process the long term recordings, an unsupervised learning scheme based on incremental local tangent space alignment （ILTSA） is chosen to cluster the new coming samples dynamically. The proposed method is evaluated by vibration signals measured on defective bearings with different fault types and pressure signals collected from compres- sor with surge fault, respectively. The results show that the scheme enables to achieve a high accuracy for condition classification with potential in identifying novel patterns from high dimensional feature space.
出处
《西安交通大学学报》
EI
CAS
CSCD
北大核心
2011年第1期64-68,136,共6页
Journal of Xi'an Jiaotong University
基金
国家自然科学基金资助项目(51075323)
北京交通大学轨道车辆结构可靠性与运用检测技术教育部工程研究中心资助项目(SROMRGV(BJTU)2010-002)
中央高校基本科研业务费专项资金资助项目
关键词
流形学习
增量式学习
状态监测
局部切空间排列
manifold learning
incremental learning
condition monitoring
local tangent spacealignment
