The invention discloses an electroencephalogram (EEG) signal recognition method based on self-training semi-supervised learning. The method comprises steps: an acquired EEG signal is firstly pretreated, and a semi-supervised extreme energy ratio algorithm is used for extracting feature vectors of marked samples and unmarked samples; the marked samples are then used for training a support vector machine (SVM) and the unmarked samples are predicted; an unmarked sample with a high confidence level and a corresponding prediction label are then selected to be added to a marked sample set, and the unmarked sample with a high confidence level is correspondingly deleted from the unmarked sample set, and cyclic iteration is then carried out; and finally, marking of all unmarked samples is completed. The training time of the user before a BCI system is used is greatly reduced, and the recognition precision and the robustness are effectively improved.