To achieve a more desirable fault diagnosis accuracy by applying multi-domain features of vibration signals, it is significative and challenging to refine the most representative and intrinsic feature components from the original high dimensional feature space. A novel dimensionality reduction method for fault diagnosis is proposed based on local Fisher discriminant analysis (LFDA) which takes both label information and local geometric structure of the high dimensional features into consideration. Multi-kernel trick is introduced into the LFDA to improve its performance in dealing with the nonlinearity of mapping high dimensional feature space into a lower one. To obtain an optimal diagnosis accuracy by the reduced features of low dimensionality, binary particle swarm optimization (BPSO) algorithm is utilized to search for the most appropriate parameters of kernels and K-nearest neighbor (kNN) recognition model. Samples with labels are used to train the optimal multi-kernel LFDA and kNN (OMKLFDA-kNN) fault diagnosis model to obtain the optimal transformation matrix. Consequently, the trained fault diagnosis model implements the recognition of machinery health condition with the most representative feature space of vibration signals. A bearing fault diagnosis experiment is conducted to verify the effectiveness of proposed diagnostic approach. Performance comparison with some other methods are investigated, and the improvement for fault diagnosis of the proposed method are confirmed in different aspects.

为了通过应用振动信号的多域特征来获得更理想的故障诊断精度，从原始的高维特征空间中提炼最具代表性和内在的特征分量具有重大意义和挑战。基于局部费舍尔判别分析（LFDA），提出了一种新的降维方法，该方法同时考虑了标签信息和高维特征的局部几何结构。LFDA引入了多内核技巧，以提高其在处理将高维特征空间映射到较低特征空间的非线性方面的性能。为了通过减少低维特征来获得最佳的诊断准确性，K近邻（kNN）识别模型。带标签的样本用于训练最佳多核LFDA和kNN（OMKLFDA-kNN）故障诊断模型，以获得最佳转换矩阵。因此，训练有素的故障诊断模型将以振动信号最具代表性的特征空间实现对机械健康状况的识别。进行了轴承故障诊断实验，以验证所提出的诊断方法的有效性。研究了与其他方法的性能比较，并从不同方面证实了该方法在故障诊断方面的改进。