Aiming at the difficulty of vibration signals feature extraction and condition identification of motor bearings, a method based on enhanced symplectic geometry mode decomposition (ESGMD) with self-organizing auto-encoder convolution network (SOAECN) was proposed. Firstly, on the basis of symplectic geometry mode decomposition (SGMD), the vibration signals of motor bearing were adaptively decomposed into initial symplectic geometric mode component (ISGMCs), and the ISGMCs were adaptively reorganized by improved condensed clustering method to obtain the clustering symplectic geometric modal components (CSGMCs). Secondly, a new comprehensive evaluation index was proposed, which was used to screen the CSGMCs that can reflect the characteristics of vibration signals and then be reconstructed. Thirdly, convolution neural network and wavelet auto-encoder were combined, the auto-encoder convolution network (AECN) was constructed and its loss function was improved on the basis of AECN and self-organizing strategy was introduced, then the SOAECN was constructed. Finally, the reconstructed vibration signals were fed into SOAECN for automatic feature learning and condition identification. Experimental results indicate that ESGMD – SOAECN reaches 98.98% of condition identification rate. The ability of condition automatic feature extraction and automatic condition identification are better than deep learning methods such as deep sparse auto-encoder, deep de-noising auto-encoder, deep belief network and so on. The results can provide a reference for the identification of motor bearing conditions.