BCI(brain-computer interface)-MI(based Motor Imagery) training has been demonstrated to effectively enhance upper limb rehabilitation in stroke patients. However, traditional MI paradigms do not adequately address the diverse needs of patients regarding various joint movements during upper limb rehabilitation. A motor classification method based on SGE (spatial group-wise enhanced) multi-layer CNN (convolutional neural network) was proposed to improve the classification accuracy of upper limb movements. Six motor imagery paradigms targeting different joints of the unilateral upper limb were designed. A spatial grouping enhancement module was introduced to generate attention factors for each spatial location, adaptively adjusting the weight of EEG (electroencephalogram) data across different frequency bands. Features were then extracted using CNN, facilitating to the classification of unilateral upper limb movements. The method was experimentally validated on both public and self-collected datasets. The results indicate that the proposed method achieves higher accuracy for multi-class classification of unilateral upper limb movements, demonstrating its potential in BCI applications. It can be concluded that the proposed method provides new technical support for upper limb rehabilitation training in stroke patients.