一种半监督邻域自适应线性局部切空间排列的故障识别方法研究

FAULT RECOGNITION METHOD RESEARCH BASED ON SEMI-SUPERVISED NEIGHBORHOOD SELF-ADAPTIVE LINERA LOCAL TANGENT SPACE ALIGNMENT

线性局部切空间排列算法(Linear local tangent space alignment,LLTSA)是能够较好应用于模式识别问题的降维方法,但由于其属于无监督的降维方法且在降维过程中只使用全局统一的邻域参数,使得在对高维数据集进行约简时,不能利用部分样本的类别标签信息且不能根据样本空间分布的变化调整邻域参数。针对上述问题,提出了一种半监督邻域自适应线性局部切空间排列算法(Semi-supervised neighborhood self-adaptive LLTSA,SSNA-LLTSA)。该算法在LLTSA的基础上,利用部分标签信息来调整样本点与点之间的距离以形成新的距离矩阵来完成邻域构建,同时根据每个数据样本点邻域的概率密度自适应地调整邻域参数,进而得到更好的降维效果。经典的三维流形、UCI典型数据集模式识别和轴承故障诊断的实验结果表明,该算法克服了LLTSA算法无监督和使用全局统一邻域参数的不足,可更有效地寻找数据的低维本质流形,提高了识别准确率,具有一定优势。

Linear local tangent space alignment（ LLTSA） is a dimensionality reduction method which is easily used to pattern recognition. However,it is an unsupervised dimensionality reduction method and only use global neighborhood parameter,when it used to high-dimensional data for dimensionality reduction,its incapacity of using part sample class label information and self-adaptive adjust neighborhood parameter while the samples space distribution changed. Aiming at the problems above,a semisupervised neighborhood self-adaptive linear local tangent space alignment（ SSNA-LLTSA） dimensionality reduction method is proposed in this paper. In SSNA-LLTSA, the distance between different points is adjusted by utilizing part class label information,thereby a new distance matrix is formed and the neighborhood is constructed through this new distance matrix. At the same time, the neighborhood parameters are self-adaptive adjusted according to probability density of each sample point neighborhood. The experiment results of classical 3D manifold,UCI datasets and bearing fault diagnosis show that the algorithm overcomes the drawbacks that the LLTSA has no supervision and the use of global unified neighborhood parameters and it is more effective to find the low dimensional nature of the data for improving the recognition accuracy and has certain superiority.