The rotor-bearing system of large rotating machinery has multiple bearings with complex vibration correlations, which significantly affect the effectiveness of intelligent diagnosis in industrial production. In this article, a new framework of fault diagnosis for the rotor with multiple bearings is proposed. The framework is composed of two parts: 1) instantaneous orbit feature fusion image construction; 2) the deep convolutional network based on transfer learning. The multivariate complex variational mode decomposition (MCVMD) is adopted to decompose the complex-valued signals of multiple bearings, which can make full use of the joint information between signals by considering the axis orbit of each bearing simultaneously. To our best knowledge, it is the first attempt of applying MCVMD to the field of fault diagnosis. Then, multiple orbit features are derived from the decomposed signals to reflect the transient state of vibration. Finally, the fusion feature images, constructed by the orbit features of multiple bearings, can exhaustively present the overall status of the rotor-bearing system. Parameter transfer is used for the deep convolutional network to solve the time-consuming training problem. The experiment and verification is carried out on three steam turbines and the pumped storage unit. The results demonstrate that the proposed method outperforms the existing approaches based on the original signal, frequency, or time-frequency features.