There is a limitation existing in classical bearing fault diagnosis thatit is required to build different target models to fitvarying working conditions. This paper proposes a spectral centroid transfer learning model, which transfers the source working condition domainto target working condition domain; the modeling cost for target working condition domainis reduced and the universality of bearing fault diagnosis modelis enhanced.Firstly, the frequency spectrum similarity measure (FSSM) value between the two working conditions domains is calculated and the source working condition domain with near distance is sorted and selected as the initial training set. Then, during iteration process, the samples whose spectral centroid meandistances are relatively far from that of the training set are removed, and the same quantity of label less samples from target working condition domain are added to training set. The iteration finishes when the spectral centroidme and istances of both the working conditions domainsare equal. Here the fault categories are determined by the outputs of two subclassifiers: The support vector machine (SVM) and the logistic regression (LR)based subclassifiers. The experiment results on Spectra Quest geared drive train show that the diagnostic performance of the proposed model is significantly better than that of nontransfer model when the rotation speed or load changes. Meanwhile, some indexes, including the number of replaced samples, the diagnostic accuracy, the FSSM index and the time consumption can be utilized to evaluate the quality of the source working condition domain.Thus,the proposed model possesses a potential valuein solving bearing fault diagnosis issue under varying working conditions.