In order to reduce the dependence of neural network on a large number of complete data in the process of mechanical equipment fault prediction and health management (PHM). To solve the problem of rolling bearing fault diagnosis with scarce data, a multi-source domain transfer learning method is proposed. The model uses one-dimensional convolution neural network, takes the original vibration signal as the input of the model, uses two different source domains data to pre-train the model, and uses the target domain data to fine-tune the pre-training model to improve the recognition accuracy of the target domain. Using the measured data of the machinery fault simulator of spectra quest and the bearing data sets of Case Western Reserve University, in the case of few fault samples in the target domain, the classification accuracy, training speed, result stability and multi-source domain effectiveness of the model are verified respectively, and the results of migration diagnosis were compared with those of CNN, TCA JDA and SVM. The results show that the model can achieve higher classification accuracy when the fault data is scarce. In the case of the number of samples in the three target domains, the classification accuracy of the multi-source domain migration method reaches 97. 71%, 96. 28% and 94. 18%, respectively. The model has fast convergence speed and good stability.