关键词: 脑机接口, SSVEP, Transformer, 滑动窗口, 光标控制

Abstract: Wearable brain-computer interfaces (BCIs) serve as an effective technology that enables brain-controlled character spelling for patients with high-level paraplegia, imposing urgent demands for high real-time performance and accuracy of the system. However, traditional canonical correlation analysis (CCA) methods exhibit insufficient feature extraction capabilities within short time windows, leading to response lag. Therefore, this paper proposes a high-speed cursor control microsystem for wearable BCIs based on the steady-state visual evoked potential (SSVEP) paradigm and a Transformer architecture. First, targeting the Biosemi ActiveTwo EEG acquisition system, the proposed system designs a differential re-referencing preprocessing pipeline based on the Cz channel. This approach effectively resolves common-mode noise interference and significantly enhances the frequency-domain features of SSVEPs. Second, an SSVEP-Transformer network is constructed, which utilizes a one-dimensional convolutional neural network (1D-CNN) to build a time-slice embedding module, realizing feature dimensionality reduction and representation for high-frequency sampled signals. Third, a self-attention mechanism is introduced to capture global correlations and harmonic frequency features across time steps. Finally, through sliding window data augmentation technology, the training samples are expanded from 96 to over 3168 windows, which significantly enhances the decoding performance of the model for 0.8 s short time window signals. The experimental results demonstrate that the proposed system achieves a recognition accuracy of 94.6% within a short time window of 0.8 s. Compared with the conventional CCA-based method, the system reduces the control latency by 60%, thereby providing a feasible technical solution and experimental foundation for the future clinical translation and engineering deployment of high-performance brain-computer interfaces.

Key words: brain-computer interface, SSVEP, transformer, sliding window, cursor control