Fliter Bank Convolution Neural Network Based on Transformer for SSVEP Classification

Deep learning is increasingly being applied in Steady-State Visual Evoked Potential (SSVEP) Brain-Computer Interface (BCI) systems, particularly models based on Convolutional Neural Networks (CNNs). This paper introduces a novel method named Conv-Trans2, which integrates the temporal characteristics of SSVEP data and channel features by utilizing the multi-head attention mechanism from the Transformer architecture to process time-domain data directly. Furthermore, to enhance harmonic learning, an extension based on a filter bank (FB-Conv-Trans2) is proposed, further improving classification accuracy. Experiments were conducted on two publicly available datasets (Dataset 1: 35 subjects, 40 targets; Dataset 2: 70 subjects, 40 targets). The results demonstrate that FB-Conv-Trans2 outperforms other baseline methods across different time windows, achieving significant improvements in both classification accuracy and information transfer rate. The study also investigates the impact of varying data lengths on inter-individual differences.