Electroencephalography (EEG) is a non-invasive technique that can be used in brain machine interface (BMI) systems to measure and record brain electrical activity. Deep learning techniques have proved superior to conventional methods in EEG-based intent decoding. However, some deep learning models have overly complex structures while ensuring the accuracy of eEg recognition, resulting in reduced training and recognition speed. In this study, we proposed a compact multi-head self-attention deep learning decoder that combined the convolutional neural network-based EEGNet decoder with the ProbSparse multi-head self-attention mechanism. Compared with traditional self-attention methods, this decoder ensures alignment dependent on both time complexity and memory usage of O ( L log L ) and it has been demonstrated to enhance the accuracy of EEG-based intent recognition. The test results on dataset 2a from BCI Competition IV showed that the EEGNet-MSD decoder performed approximately 8% better than the competition winning decoder FBCSP e NBPW, and achieved better results than the latest LSTM neural decoding method. In addition, a binary classification test was performed on the Physiobank EEG motor imagery dataset, and the results showed that the accuracy of EEGNet-MSD was approximately 4% higher than EEGNet, validating the stability of the EEGNet-MSD decoder. This study provides a new solution for enhancing the performance of EEG-based intent decoding in both accuracy and decoding speed.

Virtual Reality Motion Sickness (VRMS), which mainly caused by visual-vestibular-somatic conflict, is a risk to the health and security of users who have fun in virtual reality. To evaluate the VRMS in virtual reality experience, this paper introduces regression models based on recurrent neural network to predict the user’s VRMS severity from their EEG data. The EEG before, during and after a VRMS exposure task is collected and divided into 5 rhythms as the inputs. The Simulator Sickness Questionnaire (SSQ) is performed after each task to label the EEG data as the VRMS severity. In the proposed four regression models, four different approaches used to extract electrode-frequency features are combined with a three-layer Long Short-Term Memory (LSTM) or Bi-directional Long Short-Term Memory (BiLSTM) network used to study the temporal features. Besides, the different numbers of hidden units in LSTM and BiLSTM network are compared to choose an optimal one. The results suggest the fusion of all electrodes and rhythms information using convolutional neural network before inputting the LSTM/BiLSTM network provides the best regression performance, in which the mean MSE and R2-score are 53.6159 and 0.8683 respectively. The work introduced in this paper provides a method to assess the performance of VR productions and is an objective and direct guideline to overcome VRMS and optimize VR systems.