다중의도 예측을 위한 딥러닝 기반 뇌-컴퓨터 인터페이스에 관한 연구

Abstract

Introduction: Brain-computer interface (BCI) is a technology that controls computers or machines using brain signals. With this technology, people with various disabilities, such as neural paralysis, and spinal cord injury can control electric devices or express their thoughts by simply thinking. The BCI types can be devided based on which electroencephalography (EEG) characteristic is used to predict intentions as follows. P300-based BCI uses amplitudes of P300. P300 is a positive peak of EEG signal that occur about 300ms after a visual stimulation. SMR is based on the characteristic that the power of alpha wave (8~13Hz) or beta wave (13~30Hz) on motor cortex is changed according to the user’s imagination of body movement. Steady state visual evoked potential (SSVEP) based BCI uses a power spectrum of EEG signal measured on the visual cortex. To use the BCI system in daily life, the BCI should be able to predict various intentions such as direction of walking, text typing, and body movements. However, the previous BCI methods have a limitation that they can predict only one type of intention. In this paper, we propose a multi-functional BCI method that can predict various intentions simultaneously.

Method: To evaluate whether the proposed multi-function BCI can or cannot predict multiple intentions, different types of intentions, steady state visually evoked potential (SSVEP), sensory motor rhythm (SMR) and both of SSVEP and SMR (Multiple Intention), were predicted by one BCI model. We used EEG data measured during SSVEP and SMR paradigm, respectively, for 54 subjects. Signal processing of the multi-functional BCI is as follows; noise was removed from electroencephalography (EEG) data using common average reference (CAR) and a band-pass filtering. After that, features that reflect user's multiple intentions are extracted using power spectrum analysis and normalization process. Finally, artificial neural networks or convolutional neural network predict multiple intentions.

Results: The prediction accuracy using convolutional neural network was 38.89% for Multiple Intentions. Chance level of the prediction was 1.56%. These results indicate that the proposed multi-functional BCI can predict multiple intentions. It also means that users of the proposed BCI system can control various electric devices simultaneously. Also, this will enable the application of practical BCI system for daily life in the near future.