Research On EEG Signal Recognition Method Oriented To Virtual Reality Target Detection

In the field of brain-computer interface,the use of EEG signals for target detection has become the focus of attention for many scholars because of its ability to directly respond to the cognitive processing of the human brain in response to external changes.The keys to using EEG signals for target detection are:(1)the analytical processing of EEG signals.In the analysis and processing of EEG signals,a small sample of EEG data sets will often cause the model to fail to learn features with strong generalization ability,and in the EEG experiment of binary classification,the categories of the dataset are unbalanced,the number of positive samples is much smaller than the number of negative samples,which causes the model to be heavily biased towards the categories with a large number of samples during training,which is very ineffective for the actual target detection problem;(2)designing an effective target detection paradigm.The traditional EEG paradigm mainly presents two-dimensional target stimuli to subjects through an ordinary monitor.Compared with the three-dimensional targets we see in real world,such two-dimensional stimuli lack realism,and it is also difficult to bring sufficient immersion to subjects in the experiment,thus making it difficult to induce a truly effective EEG signal.To address the above problems,the research in this thesis mainly includes the following aspects.(1)A feature combination and optimization method based on convolutional neural network is proposed.The traditional standard convolution is replaced by using Depthwise Separable Convolution,and multiple convolution kernels of different sizes are convolved in parallel,which greatly reduces the number of training parameters of the model.At the same time,because the Depthwise Separable Convolution separates the channel convolution and the feature map convolution,and adopts the step-by-step convolution method to adapt to the characteristics of the EEG signal,it can extract the temporal and spatial features with stronger characterization.On this basis,the features are further optimized for combination,and a plurality of balanced data subsets are constructed on the basis of the original data sets,and the neural network is used as the feature extractor,which is trained with the data subsets and then classified by the xgboost model after conflation-optimization of the extracted features.It is demonstrated that the algorithm achieves an average accuracy of 81%on the pairwise public dataset and 90% on the experimental dataset under the VR paradigm,which is an average 9% improvement in accuracy compared to the performance of other comparison algorithms on this dataset.(2)The impact of VR technology on EEG experiments was investigated,and the related experimental results were analyzed.The results show that the experimental accuracy under the VR paradigm is 1%～2% higher than that under the non-VR paradigm,both within-subject and across-subject dataset,and that the length of data required under the VR paradigm was less than that under the non-VR paradigm for the same experimental effect when segmenting the EEG signal data,which shows that due to the immersive characteristics of VR technology,so compared with the ordinary EEG experimental paradigm under non-VR,it can evoke better EEG signals in a short period of time,which helps to improve the performance of the brain-computer interface system.(3)A set of brain-computer interface system for online detection of virtual reality targets is built,which solves the problems of interaction between software and hardware,online target detection,communication with external devices,etc.,and realizes a system for online target detection in a virtual environment.In summary,this thesis solves the problems of overfitting and category imbalance in training small sample EEG data sets by using the feature combination and optimization method based on convolutional neural networks,and verifies the improvement effect of VR technology on brain-computer interface performance.Finally,a set of brain-computer interface system for virtual reality target detection is realized.