Abstract: Considering that the traditional rolling bearing fault identification methods heavily rely on expert experience and are diffcult in fault feature extraction and seletion,a method based on morphological empirical wavelet transform(MEWT)and improved FractalNet(IFractalNet)is proposed.The raw vibration signals of rolling bearings are adaptively decomposed into severalintrinsic modal components by MEWT.The components with obvious fault characteristics are selected using the comprehensiveevaluation index and then reconstructed.Considering that the defects of the raw FractalNet,the loss function,activation functionand optimization method are improved.The reconstructed vibration signals are fed into IFractalNet for automatic feature extractionand fault identification.The experimental results indicate that the proposed method avoids the complex manual feature extractionprocess and can effectively identify the various fault types and multiple fault severities of roling bearings,which shows obvious su-periorities in generalization ability,feature extraction ability,and fault identification ability.