Abstract: To address the problem of accurately extracting features from the non-stationary and not high signal-to-noise ratio vibration signals of rolling bearings, this study proposes a fault diagnosis method with high recognition rate and noise resistance. Firstly, the symplectic geometry mode decomposition (SGMD) is decomposed the fault signal into multiple symplectic geometry components (SGCs). Based on the correlation principle, the SGCs with high correlation is selected to reconstruct the fault signal, forming the reconstruction signals. Then, composite multi-scale attention entropy (CMAE) is proposed to quantitatively extract the feature values of the reconstructed signals, and construct CMAE features. The whale optimization algorithm (WOA) is used to optimize the relevant parameters in the extreme learning machine (ELM) to construct the WOA-ELM model. Finally, the CMAE features are input into the WOA-ELM model to achieve fault diagnosis of rolling bearings. The simulation and experiment results that compared to other methods,the proposed SGMD-CMAE and WOA-ELM methods have higher accuracy in identifying rolling bearing faults.