Abstract:

As a classic manifold dimensionality reduction algorithm， Locally Linear Embedding （LLE） algorithm is widely used in the field of fault diagnosis because of its good feature extraction ability. However， due to the inherent defects of LLE algorithm， such as sensitivity to neighborhood parameters and single structure of mining， the features extracted by it have poor discrimination ability in practical applications. Therefore， a Discriminant Similarity and Tangent Adaptive Neighborhood Locally Linear Embedding （DSTANLLE） algorithm was proposed and applied to the bearing fault diagnosis. Firstly， a new metric method， which fuses tangent space， was used to evaluate the local similarity between samples. Then， an adaptive neighborhood graph was constructed to select neighbors for each sample point. Finally， the discriminant similarity information was added to extract the discriminant structure of the data. Experimental results on two synthetic datasets and two bearing fault datasets show that DSTANLLE algorithm can extract significant discriminative features from data， achieving an OA （Overall Accuracy） of up to 100% in the application of bearing fault diagnosis.

Key words: Local Linear Embedding (LLE) algorithm, feature extraction, dimension reduction, tangent space metric, adaptive neighborhood, fault diagnosis

摘要：

局部线性嵌入（LLE）算法是一种经典的流形降维算法，具有良好的特征提取能力，在故障诊断领域应用广泛。然而，LLE算法固有的缺陷例如对邻域参数选择敏感、挖掘的结构单一等问题，使得它在实际应用中提取的特征存在判别能力较差的问题。为此，提出判别相似性和切空间自适应邻域的局部线性嵌入（DSTANLLE）算法，并将它用于轴承故障诊断。首先使用融合切空间的新度量方式评估样本之间的局部相似性，其次构造自适应邻域图为每个样本点选择邻居，最后加入判别相似信息以提取数据的判别结构。在2个人工合成数据集和2个轴承故障数据集上的实验结果表明，DSTANLLE算法可以提取数据中区分性显著的特征，且在轴承故障诊断应用中的总体识别精度（OA）最高可达100%。

关键词: 局部线性嵌入算法, 特征提取, 降维, 切空间度量, 自适应邻域, 故障诊断