# A time-frequency feature fusion-based deep learning network for SSVEP frequency recognition

## TL;DR

This paper introduces a deep learning network that improves SSVEP frequency recognition by combining time and frequency domain features, enhancing performance in brain-computer interfaces.

## Contribution

A novel deep learning network, SSVEP-TFFNet, is proposed that dynamically fuses time and frequency domain features to improve cross-subject generalization in SSVEP-based BCIs.

## Key findings

- SSVEP-TFFNet achieved 89.72% average classification accuracy on a 12-class dataset, outperforming existing methods.
- The model showed significant improvements on 40-class datasets with 72.11% and 82.50% average classification accuracies.
- Dynamic time-frequency feature fusion was validated as effective for calibration-free SSVEP-based BCI systems.

## Abstract

Steady-state visual evoked potential (SSVEP) has emerged as a pivotal branch in brain-computer interfaces (BCIs) due to its high signal-to-noise ratio (SNR) and elevated information transfer rate (ITR). However, substantial inter-subject variability in electroencephalographic (EEG) signals poses a significant challenge to current SSVEP frequency recognition. In particular, it is difficult to achieve high cross-subject classification accuracy in calibration-free scenarios, and the classification performance heavily depends on extensive calibration data.

To mitigate the reliance on large calibration datasets and enhance cross-subject generalization, we propose SSVEP time-frequency fusion network (SSVEP-TFFNet), an improved deep learning network fusing time-domain and frequency-domain features dynamically. The network comprises two parallel branches: a time-domain branch that ingests raw EEG signals and a frequency-domain branch that processes complex-spectrum features. The two branches extract the time-domain and frequency-domain features, respectively. Subsequently, these features are fused via a dynamic weighting mechanism and input to the classifier. This fusion strategy strengthens the feature expression ability and generalization across different subjects.

Cross-subject classification was conducted on publicly available 12-class and 40-class SSVEP datasets. We also compared SSVEP-TFFNet with traditional approaches and principal deep learning methods. Results demonstrate that SSVEP-TFFNet achieves an average classification accuracy of 89.72% on the 12-class dataset, surpassing the best baseline method by 1.83%. SSVEP-TFFNet achieves average classification accuracies of 72.11 and 82.50% (40-class datasets), outperforming the best controlled method by 7.40 and 6.89% separately.

The performance validates the efficacy of dynamic time-frequency feature fusion and our proposed method provides a new paradigm for calibration-free SSVEP-based BCI systems.