基于一种流形学习新算法的转子故障数据集降维研究

A New Manifold Learning Algorithm and Its Application on Rotor's Fault Data Set

针对旋转机械的转子故障诊断中原始特征多、难以有效提取振动信号非线性特性的问题,提出一种结合已知故障类别信息进行流形学习降维的新算法—监督邻域保持多项式嵌入(Supervised Neighborhood Preserving Polynomial Embedding,S-NPPE)。利用样本点的故障类别信息,改进转子故障特征数据集中样本点间的欧式距离,重新构造样本点邻接图,再对数据集进行非线性降维处理。先介绍流形学习降维理论,然后给出重构邻接图以及S-NPPE算法的基本步骤,结合转子试验台的不同状态下的振动信号,探讨其在转子故障数据集降维中的应用。

To solve the problem that rotor＇s original features have high dimensionalities, and it is difficult to extract non- linear features in fault diagnosis of rotating machinery. A new manifold learning algorithm called Supervised Neighborhood Preserving Polynomial Embedding （S-NPPE）IS proposed, which utilizes labels of each sample for dimensionality reduction. With fault labels, the neighborhood graph with recalculated Euclidean distances between each data points is improved. Dimensionality reduction is implemented with reconstructed adjacency graph, and then non-linear dimensionality reduction is operated. Firstly, it gives knowledge about manifold learning for dimensionality reduction. Then it presents reconstruction for adjacency graph and the basic steps of whole algorithm Finally, the S-NPPE algorithm is used on rotor＇s fault data set with different fault classes.