基于LLTSA的转子故障数据集降维方法

Dimension Reduction Method for Rotor Fault Dataset Based on LLTSA

特征数据集降维是机械故障智能诊断的关键步骤之一。常见的数据集降维方法难以准确的从高维非线性数据集中获取反映转子运行状态的敏感特征信息,导致故障模式识别精度降低。局部切空间排列算法(LTSA)对部分高维非线性数据集达到较好降维效果。但该算法不适合处理高曲率分布、稀疏不均匀分布等高维数据源。为此,在LTSA算法的基础上结合线性分块思想,提出线性局部切空间排列算法(LLTSA)。该算法充分考虑了数据集的整体与局部结构,将数据样本空间划分为一组最大线性块,使降维后的同类数据具有更好的聚类性。通过高维非线性转子振动数据时域特征数据集对该算法进行验证,结果表明经该算法降维后的数据集具有较好的聚类与分类性能。

Characteristic dataset dimension reduction is one of the key steps in mechanical fault intelligent diagnosis. However, it is difficult for the conventional dataset dimension reduction method to accurately extract the sensitive characteristic information, which can reflect the operation state of the rotor, from the high-dimensional nonlinear datasets. Thus, the precision of fault pattern recognition will be reduced. Although the local tangential space alignment （LTSA） algorithm can yield good dimension reduction effect for a part of the high-dimensional nonlinear datasets, it is not suitable for processing the high dimensional data sources with high-curvature distribution, sparse uneven distribution etc. In this study, combining the LTSA algorithm with the linear blocking method, the linear local tangential space alignment （LLTSA） algorithm is proposed. In the algorithm, the global and the local structures of the dataset are fully considered, and the data sample space is divided into a set of linear blocks so that the same data after the dimension reduction has a better clustering performance. This algorithm is verified by the time-domain characteristic dataset of the high-dimensional nonlinear rotor vibration data. The results show that the dataset after the dimension reduction algorithm has a good clustering and classification performance.