This paper proposes a novel method based on hyper-ellipsoidal Support Vector Data Description (SVDD) for bearing fault diagnosis. First, features of bearing fault data are selected based on integrated indicators to solve the overlapping problems of features from different bearing faults. Second, considering that multiple fault data of bearings in practical applications cannot be obtained at one time in a short time, the incremental learning model is established by creating high-dimensional spatial feature hyper-ellipsoids with the concept of SVDD. Finally, we conducted experiments by two laboratory data sets to validate the effectiveness of the proposed method.