Research And Application Of SSVEP Feature Extraction And Classification Algorithm

Brain-computer interface(BCI)constructs an information pathway for direct interaction between the brain and external devices,which can help patients with motor dysfunction to restore certain communication and locomotion capabilities.This thesis is dedicated to the research of brain-computer interface based on steady-state visual evoked potential(SSVEP),which has the advantages of high information transfer rate and little training,therefore it has a wide application prospect.Although methods based on canonical correlation analysis(CCA)have high classification accuracy,background noise and individual differences still have a great impact on the stability of the system,and the recognition accuracy of short time window signals needs to be further improved.This paper conducts research on the above issues,and the main results are summarized as follows:Firstly,a SSVEP-based brain-computer interface system is designed in this paper,including the hardware implementation of the visual stimulus device and the design of the stimulus paradigm.The system can generate precise stimulus frequencies and has real-time signal processing capability.And based on this system,a wheelchair robot navigation system that can be controlled by EEG is built through simulation software.Secondly,in order to reduce the adverse effects of individual differences,and solve the problem that traditional CCA method does not make full use of the information in harmonic frequency band,this paper proposes a multiple sub-band frequency recognition algorithm based on individual template signals(Mset FBCCA).The algorithm extracts common features from training samples to construct individual template signals,thereby reducing the impact of individual differences.In addition,the test signal is decomposed into multiple sub-band components through the filter bank,and then analyzed independently to obtain the characteristics,which efficiently utilizes the information in multiple harmonic frequency bands and reduces the mutual interference among them.Experimental results show that Mset FBCCA has better recognition performance under time window greater than 1.2 seconds,and the accuracy is 13.56±4.25%higher than CCA,and 3.30±1.14% higher than Mset CCA.Thirdly,considering that Mset CCA is likely to identify background noise as common features,this paper proposes MCM-FBCCA algorithm as an extension of Mset FBCCA by combining multilayer correlation maximization(MCM)model.The algorithm selects samples that are highly correlated with the stimulus frequency for training,and uses the typical variables to carry out multi-level joint spatial filtering,which effectively improves the SNR of the template signals and further improves the recognition accuracy of SSVEP.Experimental results show that MCM-FBCCA can accurately extract the characteristics related to the stimulus frequency from the training samples to construct the template signals.Compared with other algorithms in this paper,MCM-FBCCA has the best recognition performance and is least affected by individual differences.Under the short time window of 1.2 seconds,the accuracy of the optimized MCMFBCCA is 89.0±7.98%,which is 32.17% higher than CCA and 10.08% higher than Mset CCA.Finally,to solve the problem of large uncertainty in the brain-controlled wheelchair navigation system,the proposed MCM-FBCCA algorithm was combined with the artificial potential field(APF)path planning method to achieve shared control,and 5 subjects were recruited for an online interaction experiment.The experimental results show that the online accuracy of the system reaches 82.99±10.99%,which is 14.11% higher than that of CCA,and the stability of the system is significantly improved.