The invention discloses a multi-source-domain self-adaptive cross-subject EEG cognitive state evaluation method based on label alignment. The method comprises the following steps: 1, data acquiring; 2, data preprocessing; 3, a cross-subject EEG cognitive state evaluation method based on the LA-MSDA model. According to the method, a shared common feature extractor and a non-shared sub-feature extractor are used in stages, and tested invariant features and specific features of a source domain sample and a target domain sample are further learned; in consideration of the relationship and similarity between cross-subjects, a method for aligning inter-domain distribution of local and global representation is provided to evaluate the cognitive state of the cross-subjects, and the problem that it is difficult to learn fine-grained class condition information and adapt to decision boundary samples of the cross-subjects is solved. Finally, the problem of individual difference of electroencephalogram signals in the field of brain cognitive calculation is effectively avoided, the method can be suitable for cognitive state recognition based on EEG under any task, the generalization ability is high, and the method can be well suitable for clinical diagnosis and practical application.