An Improved Multidimensional Filter Bank Canonical Correlation Analysis for Recognition of SSVEP-Based BCIs

Abstract

This letter presents an improved multidimensional filter bank canonical correlation analysis (FBCCA) method for the brain-computer interface (BCI) system based on steady-state visual evoked potentials (SSVEP). This is a training-free SSVEP recognition method based on FBCCA, which integrates partial least squares regression (PLSR) and adaptive multidimensional extension (AME). Compared to FBCCA, this new method can further eliminate noise and artifacts from EEG signals during dimensionality reduction and regression by minimizing distribution errors. Additionally, it more effectively utilizes the valuable information from multi-channel EEG signals, thereby enhancing the recognition performance of SSVEP. Offline experiments conducted on two different open-source datasets verified that this method achieves advanced performance in training-free methods across different gaze times. In online tests on a real-time eight-target BCI system, the method achieved a peak accuracy of 98.44% and an information transfer rate (ITR) of 45.68 bits/min. This method improves the accuracy and efficiency of training-free SSVEP recognition, facilitating the wider application of BCI systems in real-life scenarios.