Abstract
A motor imagery (MI)-based brain–computer interface (BCI) is a promising system that can help neuromuscular injury patients recover or replace their motor abilities. Currently, before one uses MI-BCI, we need to collect a large amount of training data to train the decoding model, and this process is time consuming. When trained with a small amount of data, existing decoding methods generally do not perform well in MI decoding tasks. Therefore, it is important to improve the decoding performance with short calibration data. In this study, we propose a dynamic weighted filter bank domain adaptation framework that uses data from an existing subject to reduce the requirement of data from the new subject. A filter bank is used to explore information from different frequency subbands. A feature extractor with two 1-D convolutional layers is designed to extract EEG features. The class-specific Wasserstein generative adversarial network (WGAN)-based domain adaptation network aligns the distribution of each class between the data from the new subject and the data from the existing subject. Additionally, we apply an attention network to dynamically allocate different weights for different frequency bands. We evaluate our method on a public MI dataset and a self-collected dataset. The experimental results show that the proposed method achieves the best decoding accuracy among the compared methods with different amounts of training data. On the public dataset, our method achieves 8.88% and 7.16% higher decoding accuracy than the best comparing method with on block of training data on the two sessions, respectively. This indicates that our method can enhance MI decoding accuracy with a small amount of training data.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have