In order to diagnose the fault of planetary gear, the collected signals were decomposed by using complete ensemble empirical model decomposition with adaptive noise(CEEMDAN) method. The correlation coefficient of each IMF component obtained by decomposition was calculated, and the first 4 order components with greater correlation with the original signal were selected for sample entropy calculation, and the eigenvalues were obtained to form the eigenvectors. The feature vectors were put into the probabilistic neural network system for diagnosis and compared with the diagnosis results based on the local mean decomposition sample entropy feature extraction method. The results show that by using the CEEMDAN sample entropy feature extraction method, the vibration characteristics of the system can be more accurately reflected, and the fault diagnosis accuracy is higher.