出优化改进变分模态分解方法（WOA-IVMD）将轴承振动信号分解至不同频段；又考虑信号非线性，通过9种非线性特征参数，基于经WOA-IVMD分解分量构建非线性“复合高维”特征矩阵，为避免高维数据导致维数灾难问题，采用随机近邻嵌入理论（t-SNE）对高维特征矩阵进行降维处理，并以降维所获数据作为测试样本，通过神经网络完成轴承工作状态分类。结果表明：WOA-IVMD分解信号具有与原分量更高的相似度；采用t-SNE对非线性“复合高维”矩阵进行降维，其三维流形表现具有突出的分类效果；以降维数据为测试样本，采用神经网络进行学习建模并分类，其结果具有较高的吻合度，表明提出方法可准确进行轴承状态分类。

The wind turbine bearing fault signals are lacks of characteristics for describing the bearing working states,and its nonlinearity increases the difficulty of fault diagnosis of wind turbine bearings.The Improved variational mode decomposition method is proposed in this paper and named as WOAIVMD which considers the interaction of the preset parameters,and uses the Whale Optimized Algorithm to identify the optimized parameters.Fault signals are decomposed into different modes in frequency domains by WOAIVMD,and high dimension feature matrixes are built through 9 nonlinear characteristics and all modes.In order to deal with the problem of dimensionality,tSNE method is used to reduce the dimension of the nonlinear characteristic matrix.And the neural network is use to classify different working states through using the reduction databases as a sample to build a fault diagnosis model.The results show there is good similarity with original signal in decomposing signal by proposed WOAIVMD.There is good performance in manifold with three dimensions,and it would accomplish fault classification by tSNE.In the meantime,a fault diagnosis model built by neural network has high accuracy in fault diagnosis.

X511

国家自然科学基金（51976131,51676131,51176129,51875361）