A Dual Training Framework Addressing Channel and Frequency Discrepancies in SSVEP Classification

Various multichannel frequency recognition methods have been developed to enhance steady-state visual evoked potential (SSVEP) decoding performance. However, only a few methods can effectively address discrepancies in SSVEP responses across channels and frequency bands. Therefore, this study proposed multivariate variational mode decomposition-informed ensemble task-related component analysis method (MVMD_eTRCA), a dual training method addressing the SSVEP discrepancies across channels and frequency bands. Specifically, a set of weights is trained by using the dung beetle optimizer (DBO) for components extracted via MVMD, enhancing relevant SSVEP components to amplify the desired elements and attenuating others to reduce interference, thus improving the signal-to-noise ratio (SNR). The weighted reconstruction signals serve as individual templates, and the eTRCA is applied to train a spatial filter that maximally enhances the SNR of the data, enabling target recognition. Evaluation and comparison using SSVEP-BCI public datasets show that the recognition accuracy of MVMD_eTRCA is about 68% higher than MVMD_canonical correlation analysis (CCA), approximately 12.6% higher than FBeCCA, and about 6.2% higher than eTRCA. These findings indicate that the proposed method can achieve high recognition accuracy in SSVEP mode decoding, particularly with short data lengths, potentially improving information transfer rates and reducing training time. The advantages of our proposed method demonstrate great potential for developing and improving brain-computer interface (BCI) systems.