Abstract: Aiming at the defect that the original LPP ( Local Preserving Projection) algorithm is difficult to accurately obtain the local manifold structure of non-uniform high-dimensional data and can not use the sample category information, a MIF-LPP (Multi-Information Fusion Local Preserving Projection) algorithm is proposed. MIF-LPP algorithm uses the improved standard Euclidean distance to obtain the nearest neighbor information and mutual neighbor information of samples, reducing the impact of uneven distribution of sample points and the difference of data dimensions of a single sample. The weight matrix is constructed by fusing the class information of the samples, and then the low dimensional essential manifold of the data is obtained. The validity of the algorithm is verified on CWRU(Case Western Reserve University) data set and our laboratory bearing data set respectively. The experimental results show that the feature extraction performance of MIF-LPP algorithm is obviously superior to other algorithms, and it is robust to neighborhood values. 

Key words: local preserving projection, standard Euclidean distance, multi-information fusion, bearing fault diagnosis