Brain-Computer Interface of Emotion and Motor Imagery Using 2D Convolutional Neural Network — Recurrent Neural Network

Abstract

Brain-Computer Interface (BCI) allows human communication without gestures or muscle movements with external devices. BCI works by translating commands from the brain into actions through intermediate devices such as Electroencephalogram (EEG). Usually, EEG captures when imagining movement, which is called motor imagery. Some BCIs consider other variables of the EEG signal to execute commands, such as emotion or attention. Each variable has different characteristics and is recorded from different channels, so it is necessary to use multi networks. The EEG signal comes from multiple channels. Therefore, it needs to be connected between channels besides time series. This study proposed the identification of BCI using multiple networks. Each network is identified using a two-dimension Convolutional Neural Network (2D CNN) - Recurrent Neural Network (RNN). The EEG signal is pre-filtered using a Wavelet containing Motor Imagery and two emotion states: happy and neutral. The results showed that using multi networks increased the accuracy by 93.55% compared to 83.87% for using a single network. The difference in accuracy is more significant than the use of 2DCNN-RNN, which increased by 3% against RNN alone and only 1.5% against 1DCNN-RNN. Experiments show that the LSTM memory performed almost the same as the GRU. The experiment also tested the learning parameters, which showed a small learning rate of 0.000350 providing high accuracy and short transient time.