An improved adaptive noise-complete ensemble empirical modal decomposition (CEEMDAN) with correlation dimension for petrochemical bearing fault feature extraction was proposed for the problem that the bearing vibration signals of petrochemical units are difficult to be distinguished automatically.The bearing vibration signals under four working conditions in the bearing fault data measured by a key laboratory of fault diagnosis were selected and tested for analysis.The measured vibration signals were decomposed by using the modified CEEMDAN to obtain multiple modal components IMF.The selected components were superimposed and summed to obtain the embedding dimension and delay time of the data and phase space reconstruction was performed.Combined with G-P algorithm,the correlation dimension under different embedding dimensions was found for feature extraction.And the experimental accuracy is 92.5% by the extreme learning machine,which proves the effectiveness of the method.