Multi-classification based on motor imagery EEG signals currently suffers from the problems of slow transmission rate and low classification accuracy. In this paper, we study four classes of multi-channel motor imagery EEG signals. Firstly, we proposed a classification method based on “One-vs-Rest”CSP feature extraction and sparse embedding, by extracting the CSP features of the training data through “One-vs-Rest” feature extraction, embedding the corresponding label matrix in low dimensions to obtain the embedding matrix of the training data, and then learning and training a regression model. After feature extraction of the test data, the regression model calculates the embedding matrix of the test data, and finally performs KNN classification of the training data embedding and the test data embedding in the embedding space. This method reduces the complexity of feature extraction for multi-class tasks and improves the classification efficiency, the BCI Competition 2008IV-2a public data-set is used for experimental testing and comparison with alternative classification techniques to confirm the algorithm's efficacy. In comparison to other EEG signal classification algorithms, the experimental findings show that the strategy suggested in this study achieves a certain improvement in accuracy and a shorter running time.