The original data set of rolling bearing contains high-dimensional non-sensitive features，so a low-dimensional sensitive feature extraction method through integrating kernel principal component analysis（KPCA）and t-distributed stochastic neighbor embedding（t-SNE）is proposed in this paper. The time domain，frequency domain and time-frequency domain features of the raw rolling bearing vibration signal are calculated to construct the original high-dimensional feature data set. KPCA is used to reduce the correlation of high-dimensional data set，and the nonlinear feature subset is extracted while maximizing the global feature variance of the data set. t-SNE is employed to mine the local structure information of feature data set，and further obtain the low-dimensional sensitive feature subset with high discriminability. The low-dimensional feature subset is input into the k-nearest neighbor classifier（KNNC），and the classification accuracy and clustering results are used as the quantitative indexes to evaluate the performance of the feature extraction method. In this process，the global and local structure features of the data set are comprehensively considered and the structure information of the data itself is fully utilized，so the low-dimensional sensitive features can be accurately extracted. The proposed method is applied to the fault diagnosis of a rolling bearing test rig. By comparing the results obtained from 6 different feature extraction methods and analyzing the fault data sets under variable rotating speeds and different random noises respectively，the effectiveness of the proposed method is verified，and consequently it is actually an excellent feature extraction method for the fault recognition of rolling bearings.