Abstract: In the application of kernel principal component analysis, cumulative contribution rate method is used to determine the number of kernel principal component usually, which abandon some kernel principal components whose contribution rate is small. It loses part information of samples and influences fault diagnosis effect. Aiming at this fact, a kernel principal component analysis method based on class mean is proposed. After data samples in input space are mapped into higher-dimensional space, class mean vectors of mapped data are determined, and then the PCA method is used to analyze the class mean vectors in the subspace of class mean vectors. Construct class mean kernel matrix, and make use of it to construct algorithm of class mean kernel principal component. The feature vectors formed by class mean kernel principal component include all variable information of initial data and its dimension is lower than the number of fault category. It can realize dimensionality reduction without information loss based on class mean vector. The improved algorithm is applied to rolling bearing fault diagnosis, and the results show that it has the stronger ability of integrating original variable information than KPCA, which can extract classified information of data samples more effectively and recognize fault accurately

Key words: kernel principal component analysis;class mean kernel principal component analysis;fault diagnosis