The study of fault feature extraction and early warning of rolling bearing vibration signal of generator sets is beneficial for the timely diagnosis of bearing faults, thus improving the service life of generators. In this paper, a combined EEMD-GRU-MC prediction method is adopted to predict the model based on GRU through the data decomposition of EEMD, and the predicted model residuals are corrected using MC. The analysis and diagnosis of the algorithmic model are used to determine the fault characteristics of the generator’s vibration signals for diagnosis, and the analysis and diagnosis of the characteristics are verified using experiments with publicly available data sets from the Bearing Data Center at the Paderborn University School of Mechanical Engineering in Paderborn, Germany. Diagnosis can be performed with an accuracy of 99.6% under condition load