Fault intelligent diagnosis of rolling bearing is mostly based on the same data source. Due to the difference of time and frequency characteristics and the influence of background noise， the accuracy of identifying different models and different working conditions of rolling bearing is low. In order to solve this problem， two kinds of deep groove ball bearings， 6307 and 6205， are taken as the research objects in this paper. A transfer diagnosis model with deep belief network （DBN） as the core is established， and the characteristic identification parameters represented by waveform index， kurtosis index， approximate entropy and dispersion entropy are constructed. The maximum correlated kurtosis deconvolution （MCKD） method is introduced to suppress the influence of signal transmission path （resonance band difference） and background noise. Results show that the accuracy of the transfer diagnosis model， which is composed of MCKD and DBN， is more than 95% among the three different kinds of data， providing a feasible way for the transfer diagnosis of rolling bearing.