Preprocessing: After bandpass filtering with a pass band of [0.5~8] Hz, we perform eye movement artifacts removal using independent component analysis (ICA) for the whole dataset. SVM Training and Classification: Downsample the filtered signal with a factor of 5. Then train the SVM using data from a subset of channels with a time window of 100-850ms post-stimulus. Here the channels used for subject A and subject B are different. Channels Fp1, Fp2, Fz, Fc1, Fc2, C3, Cz, C4, Pz, PO7 and PO8 are used for Subject A, and channels Fz, C3, Cz, C4, P3, Pz, P4, PO7, PO8, and Oz for Subject B. After obtaining the optimal SVM parameters through cross-validation in the training process, we classify the samples in the test set and obtain the predicted chars. Bagging: Bagging method is used to boost the performance of the SVM classifier, and final classification results are obtained through the voting of multiple SVM classifiers.