Aiming at the nonlinear and nonstationary characteristics of the reciprocating compressor vibration acceleration signal, a reciprocating compressor bearing fault feature extraction method was proposed based on complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) and refined composite multiscale dispersion entropy (RCMDE). When using the CEEMDAN method to decompose the signal, through different parameter combinations, different IMF components could be obtained; the kurtosis values of the reconstructed signal were calculated under different parameters, the set of parameters with the largest kurtosis value was selected to carry out CEEMDAN decomposition for the signal again, and the signal reconstruction was carry out. The reconstructed signals were analyzed through RCMDE, the fault feature vectors were extracted, and the support vector machine (SVM) was used for classification and recognition. The CEEMDAN decomposition method with optimized parameters was compared with the original CEEMDAN decomposition method. The results show that by using the CEEMDAN decomposition method with optimized parameters, the periodic impact signal of the reciprocating compressor can be better extracted, which is beneficial to improve the accuracy of fault diagnosis.