针对故障特征数据维数高、非线性且系统难以建立物理模型的故障诊断问题,提出了一种全局的无关线性图嵌入故障特征提取算法.通过监督学习建立原始特征的关系图,以线性图嵌入为框架进行特征降维.特征的降维过程既保留了同类数据的局部结构,又考虑了异类数据之间的全局分布,同时最大程度地消除了特征之间的统计相关性.在标准故障数据集上的实验结果表明:与已有的经典算法相比,能更有效地提取出故障的典型特征,因而更有利于故障诊断系统训练网络的快速收敛,实现快速、准确的故障诊断. Aiming at the problem of fault diagnosis with high dimensionality and nonlinearity of fault feature data and difficult to establish physical model in system, a global algorithm for feature extraction based on irrelevant linear graph embedding fault is proposed. By constructing a relational graph of original features by supervised learning, The feature dimensionality reduction process not only retains the local structure of the same kind of data, but also considers the global distribution between the heterogeneous data, and at the same time eliminates the statistical correlation between the features to the maximum extent.In the standard fault data The experimental results on the set show that compared with the existing classical algorithms, the typical features of the fault can be extracted more effectively, which is more conducive to the rapid convergence of the training network of the fault diagnosis system and fast and accurate fault diagnosis.