A BCI System with Motor Imagery Based on Bidirectional Long-Short Term Memory

Abstract

A bidirectional long short term memory (BiLSTM) neural network was embedded into a brain-computer interface (BCI) system based on motor-imagery (MI) in this paper. The MI-based electroencephalogram (EEG) signals were used to recognize different imagery actions. The dynamic characteristics of MI signals in EEG are usually low signal-to-noise ratio as non-stationary time series. A lot of strategies have been proposed to clustering MI-EEG signals. However they are not considering the concept of series features of the signal in time domain with forward and backward manners, so the recognition results are not promising. The discrete wavelet transform (DWT) was also used to get the frequency feature from transforming each channel of MI-EEG in this paper. Then the proposed BiLSTM is proposed as a classifying system to identify the MI-EEG data. BiLSTM can extract dependencies of different time points by each recurrent unit with an adaptive manner. Besides the forward manner of time series signals in the LSTM unit, the BiLSTM also puts the output signals into previous layers with backward manner. The BiLSTM system can get more promising results in the classification of MI-EEG than those obtained by other strategies shown as in experimental results.