A Novel Unsupervised Frequency Recognition Method for SSVEP-Based BCIs Using Adaptive Updating and Transfer Learning Strategy

The application of supervised recognition algorithms based on steady-state visual evoked potentials (SSVEP) has gained increasing attention in the field of signal processing. However, these methods typically demand a large amount of training data to ensure accurate recognition, resulting in high training costs and poor human-computer interaction experiences. To address these issues, a novel method called filter bank unsupervised adaptive transfer-extended CCA has been proposed for SSVEP recognition. Unlike previous methods, this approach eliminates the need for calibration data from the current subject for training. It employs a transfer learning strategy to construct a transfer average template, utilizes an adaptive updating strategy to update the transfer template and the spatial filter with the current subject's data, and incorporates the filter bank analysis strategy to achieve high-performance unsupervised SSVEP recognition. This method was compared with three state-of-the-art unsupervised approaches, and experimental results demonstrated its superior recognition accuracy and information transfer rate. Specifically, when the time window length is 1.0 s, FBUAT-eCCA achieves an accuracy of 84.69 ± 6.33% and an information transfer rate of 148.90 ± 15.60 bits/min. Additionally, FBUAT-eCCA shows strong performance in extracting inter-class separation features.