Abstract: For a rolling bearing, especially under variable working conditions, it is difficult and even unable to obtain a large number of tagged vibration data, and then the fault diagnosis accuracy is low, a new fault diagnosis method of the rolling bearings is proposed based on variational mode decomposition（VMD） and multiple feature structure combined with transfer learning. By using VMD, each state vibration signal of the rolling bearing is decomposed into a series of intrinsic mode function（IMF） components, then the singular values of the IMF matrix can be obtained by singular value decomposition, and the singular value entropy can be calculated. The multiple feature set is constructed with the singular values, the singular value entropy, the time domain and frequency domain features. At the same time, semisupervised transfer component analysis（SSTCA） is introduced, and its kernel is constructed as the multi-kernel function. The sample features of different working conditions are all mapped to a shared reproducing kernel Hilbert space, and which can improve the compactness and interclass distinction of the data. By using the maximum mean discrepancy embedding method, more effective data are selected and regarded as the source domain. The source domain feature samples are input into support vector machine（SVM）, and then the feature samples of the mapped target domain are tested. The experimental results show that, compared with other methods, the proposed multi-kernel SSTCA-SVM method has a higher accuracy in the multi-state classification of the rolling bearings under variable working conditions.