An Adaptive EEGNet-T Distribution Decoding Algorithm Based on K-L Divergence and Deep Clustering

‍ ‍Brain-computer interface （BCI） is a pathway between the brain and the outside world that are not established through nerves or muscles. Electroencephalogram （EEG） decoding is one of the keys to affecting the performance of BCI by classifying EEG features to interpret the output brain intention. Due to the non-stationary characteristics of EEG signal， the characteristics of EEG signal will change with time even in the same experiment process， resulting in the precision of the pre-trained decoding precision model often gradually decreases with time. This study proposes an adaptive EEGNet-T distribution decoding algorithm based on K-L divergence and deep clustering （KL-Deep Clu）. The KL-Deep Clu evaluates the non-stationarity of the EEG using the K-L divergence of T distribution before and after the change of EEG characteristics and an objective function based on the change of the stationarity. And adjust the EEGNet network parameters by using this objective function to reduce the stationarity difference by changing the nonlinear mapping. Then EEGNet-T distribution is adaptively modulated according to this function to realize the adaptive decoding of EEG. Ten subjects participated in the visual-auditory BCI experiment， and conducted long-term EEG decoding prediction. Compared to the traditional algorithms， the KL-Deep Clu obtains the highest average accuracy of 87.85% （p<0.05） under 128 trials. And it showed the strongest stability in the first half of the experiment and the second half of the experiment. The stable decoding effect of the KL-Deep Clu ensures the decoding accuracy of BCI for a long time， and provides the basis for the practical application of BCI.