The Coding/Decoding Of SSVEP Brain-Computer Interface And Its Application

Nowadays,because of the rapid development of Brain-Computer Interface(BCI)technology,more and more universities and enterprises are gradually joining the team to promote the development of BCI.Therefore,in recent years,the literature and invention patents in the field of BCI have increased rapidly.To Evoked the basic needs of life for patients with physical movement disorders,Steady-state Visual Evoked Potential(SSVEP)-BCI technology came into being at the right time.SSVEP has attracted much attention from researchers because it needs short training time,easy to attract the pattern features and easy to get a high information transfer rate(ITR).Although SSVEP recognition based on Electroencephalogram(EEG)signals has achieved very good results,SSVEP recognition in a short time window is still a challenging problem.To advance in the ITR of BCI system,researchers are constantly trying to shorten the signal length required by algorithm recognition as much as possible while maintaining high accuracy.This thesis mainly carried out three aspects of work.First,a SSVEPBCI detection system was built.Then,a training-free algorithm was proposed to improve the ITR of BCI system.Last,a deep learning model was proposed,which is a training method that combines convolutional network and transfer learning and got a highly ITR for the demand of SSVEP real-time recognition.This thesis mainly describes the research work of independently building the SSVEP-BCI system,proposed the method of improving the ITR of the BCI system based on the SSVEP training-free algorithm and the training method that combines the convolutional network and transfer learning to improve the ITR of the BCI system.The main framework of this thesis is as follows:(1)The established SSVEP-BCI system was based on the hardware devices a signal amplifier and a 32-channel electrical cap from Brain Products,and based on a program that presents stimulus and a program that recognition of SSVEP signals from a toolbox Psychtoolbox of Matlab.By collecting data and using four training-free algorithms to analyze the off-line data,the test results verified the effectiveness of the system.(2)The kurtosis value in statistical theory were applied to select an appropriate kurtosis value as the threshold of SSVEP calibrated-free algorithm,and to realize the dynamic time window adjustment of SSVEP.To improve the accuracy of target recognition in the shortest possible time to achieve improvement of ITR,the length of the time window can be adjusted on the basic of the threshold value.For evaluation,the Benchmark dataset and four algorithms(Multivariate Synchronization Index(MSI),Canonical Correlation Analysis(CCA),Temporally Local Canonical Correlation Analysis(TCCA),and Filter bank Canonical Correlation Analysis(FBCCA))were applied to evaluate the recognition effect of dynamic window based on kurtosis.Experimental results showed that when the kurtosis is between 3.5and 4,the performance of average ITR could achieve the best effect,and the highest ITR could reach up to 352.90 bits/min.In addition,this method was used in the 2021 BCI Robot Contest in World Robot Conference Contest.Using the strategy of CCA combining kurtosis value for dynamic window,the average ITR of five subjects was achieved 114.94 bits/min,and our team ranked fifth in the final contest.(3)A parameter-based transfer learning-convolutional neural network(PTL-CNN)approach for SSVEP-BCI system was proposed,which can automatically fuse and extract inerand intra-subject features in EEG signals.First,we proposed a shallow CNN architecture and adopt a short time window to learn a pretrained model in a dataset with numerous subjects to explore the universal feature across subjects.Subsequently,a new subject’s calibrated data was utilized to fine-tune a final model that was unique to this subject.Experimental results demonstrated that PTL-CNN achieved a remarkable performance and significantly outperforms the compared algorithms in short time windows.As an illustration,in a time window of 0.4 s,PTL-CNN obtained an average accuracy of 80.60% with an average ITR of 247.77 bits/min on the Benchmark dataset and an average accuracy of 66.91% with an average ITR of 185.90bits/min on the Beta data.The results were significantly improved compared with those of the best compared method Ensemble-TRCA(Benchmark: 71.21%,209.12 bits/min;Beta: 53.04%,135.53 bits/min).In this thesis,self-built SSVEP-BCI system was used to study the induction mechanism of SSVEP-BCI and its recognition algorithm in the first part.In the second and third parts,based on machine learning and deep learning respectively,methods to improve the ITR of SSVEPBCI system under a short time window were proposed.After experimental testing and verification,they could achieve good expected results and provide a new idea for the algorithm research of SSVEP.