Analysis Of EEG Signals Of Chinese Imaginary Speech Based On Deep Learning

Aphasia refers to the damage of brain tissue related to language function,resulting in impaired ability of patients to understand and express communication symbol systems of human beings,which severely affects their external communication and quality of life.Based on the brain activity of internal arousal,the brain computer interface(BCI)based on imaginary speech can achieve natural communication without requiring auditory language,providing a new communication method for patients with aphasia.Therefore,in-depth research on this technology is of great significance for patients with language disorders.Currently,there are some problems in the research of imaginary speech BCI,such as difficult processing of non-stationary EEG signals,small dataset volume,strong individual differences,lack of effective features,and less research on imaginary speech of the Chinese-Tibetan language family,especially Chinese.To solve these problems,this thesis takes EEG signal analysis theory as the basis,and Chinese imaginary speech analysis as the goal,and conducts research from several aspects,such as feature extraction of imaginary speech EEG signals,user feature model establishment,imaginary speech feature embedding,and classification.A Chinese imaginary speech EEG signal analysis system is built.The specific research contents include the following aspects:(1)To solve the problems of difficulty in processing non-stationary EEG signals and small dataset volume,this thesis proposes a method of imaginary speech EEG signal analysis based on the twin neural network.This method constructs feature vectors applicable to the analysis of non-stationary signals to solve the difficulty of analyzing non-stationary EEG signals.It uses twin neural networks to extract feature embedding to solve the problem of small EEG signal dataset volume.It measures the non-similarity between different EEG signals by the Euclidean distance between feature embeddings,and realizes the category output of imaginary speech EEG signals by the adaptive k-nearest neighbor(k-NN)method.(2)To solve the problems of individual differences in EEG signals and lack of effective features,this thesis proposes a method of imaginary speech EEG signal analysis based on particle swarm optimization and twin neural networks.This method first expands the feature space from multiple perspectives,and then constructs the user’s optimal feature model by combining particle swarm optimization and twin neural networks.The optimal feature model is used to establish the user feature space and realize the analysis and category output of imaginary speech EEG signals.The multi-level extended feature space is beneficial for in-depth analysis of imaginary speech feature differences,and provides a feature analysis foundation for the analysis of Chinese and English imaginary speech EEG signals.The combination of particle swarm optimization and twin neural networks can effectively solve the problem of individual differences in EEG signals and achieve optimal analysis results for user individual feature space construction.(3)To verify the feasibility and effectiveness of the proposed methods,this thesis conducted a series of experiments using international English datasets and lab-collected Chinese datasets.The experiments verified and analyzed several aspects such as channel selection,feature extraction,user feature model establishment,deep learning for feature embedding,classification accuracy,and comparison of Chinese and English feature models.The experimental results show that the proposed method achieves good results in classifying English words and has a decent performance in analyzing Chinese vowels.In addition,through the comparison analysis of Chinese and English user feature models,it was found that instantaneous frequency features and Higuchi fractal dimension features have a tendency towards Chinese imaginary speech,which can provide theoretical and experimental references for the subsequent research on Chinese imaginary speech.