Study On Brain Signal Feature Extraction Based On The Model Of Multivariate Autoregressive

In the recognition of the signal, the most important problem is to extract the featureof signals. The classification is according to its specific characteristics of different signalextraction. In this paper, it is researched observation noise AR model parameter estimationand containing multi-channel AR model parameter estimation which is according to thetraditional AR model ideal for handling no noise and signal problems withone-dimensional.Firstly, it is studied the brain signal feature extraction based on AR model. The ARmodel with white noise and color noise are introduced in detail, and it is researedparameter estimation algorithm. And last it is applied to brain-computer interface.Combine with the method of NA_MEMD, it is applied to two groups datasets of BCICompetition Dataset, which were BCIIV_III and BCIIV_I, and then compared with ARmodel classification results, the results show that noise AR model has the better effect onthe brain signals.Secondly, it is extended feature extraction of single-channel signal to multi-channelsignal, that is to extend the traditional AR model to multivariate AR model. This paperuses the LWR algorithm of MVAR model to estimate parameters, and combined with thefeature of MPCA dimension reduction method to eliminate the redundancy. Still on thetwo sets of data BCI competition experiments, results are better than the competitionwinners or equivalent, and higher than other recent literature classification results.Finally, this paper researched white observation noise of MVAR model parameterestimation. The estimated values of LS algorithm, ILSV algorithm and modified ILSValgorithm compared with the real value, the relative error and root mean square error,which namely RE and RMSE to measure. The results show that at high SNR, modifiedILSV algorithm has the highest accuracy rate estimates, at low SNR, LS estimationperformance is significantly reduced.