Abstract: In order to improve the classification accuracy of motor imagery brain-computer interface tasks, researchers aimed to enhance the decoding accuracy of motor imagery electroencephalogram（EEG） signals. Based on the spatial distribution and multi-lead information association of EEG, the graph neural network was constructed, and a classification model of motor imagery tasks based on residual graph convolution was proposed. The residual learning was embedded into the depth graph convolutional neural network, to improve network degradation. The layered graph pooling method was added to the model, to fully extract the motor imagery EEG feature information and improve the classification accuracy. The model achieved average classification accuracy of 93.84% and 96.39% and average Kappa coefficient of 0.917 1 and 0.953 5 on two BCI competition datasets, respectively. The model can effectively improve the classification accuracy of motor imagery brain-computer interface tasks, and has good generalization ability.