Imagery Speech Classification Based on General Successive Multivariate Variational Mode Decomposition With Different Objective Functions Having Different Weights

The imagery speech (IS) is the speech that the human beings are thinking in their brain. A brain computer interface (BCI) system is employed to translate the speech thinking in the brain to the vocabulary in the dictionary stored in the computer. Since it is another form of speech, the IS is natural and easy to be used. Therefore, the BCI system based on the IS has the great potential for facilitating the subjects to communicate to the outside world particularly for the subjects suffering from the diseases such as the speech disorder disease, the locked in syndrome and the amyotrophic lateral sclerosis. Although the IS becomes more popular in the BCI system, the working principles of the IS are unknown. Hence, the classification accuracy is not high. This introduces some challenges and hinders its practical applications. In order to further promote the development of the IS based BCI system, this paper proposes a method for improving the classification performance of the IS based BCI. In particular, a general successive multivariate variational mode decomposition (general SMVMD) with different object functions having different weights method is proposed to yield a high precision and a highly efficient classification. Multi-channel EEGs are decomposed using general SMVMD, with multivariate intrinsic mode functions (MIMFs) retained if they have high correlation to the original signal or denoised via the tunable Q-factor wavelet transform (TQWT) if not, after which features are extracted for classification via a k-nearest neighbor classifier. Our proposed method is compared to the other IS classification methods and the advanced classification methods by evaluating the performance of these method using the same two datasets, including BCI2020 and KARA one. It is found that the average accuracies yielded by our proposed method for all the subjects are 65.07% and 80.51%, which are higher than those yielded by the other methods. These results demonstrate the capability of the general SMVMD and TQWT for denoising the multi-channel signals as well as the superiority of our proposed method over the other IS classification methods.