Emotion recognition based on Electroencephalography (EEG) is an important research area in affective computing. However, EEG signals usually exhibit significant individual differences, limiting the generalization ability of existing domain adaptation methods to unseen subjects. To this end, we propose a Domain-Generalized EEG Emotion Recognition framework (DG-EER) that learns domain-invariant representations for robust emotion recognition. During model training, we design a shared encoder in the source domain to capture domain-invariant features across different subjects. In the inference phase, the model only use the pre-trained encoder and a classifier to perform emotion recognition. Extensive experiments were conducted, and the proposed DG-EER achieved average classification accuracies of 75.16% and 76.83% on the valence and arousal dimensions of the DEAP dataset, respectively, and 88.05% on the SEED dataset. Moreover, our method reduces the impact of individual differences, thus enhancing its generalization ability in emotion recognition tasks.