Research On EEG Signal Feature Extraction Based On Semi-supervised And Time Series Models

EEG(Electroencephalogram,EEG)is a type of rhythmic electrical activity that is produced spontaneously by brain cell populations,it contains a lot of information related to human consciousness.Human's study on brain electrical signal has been in a state of growth,with the development of machine learning and human demand for health care,more and more countries began to pay attention to developing the Brain Project that use the brain as a research object.Through extracting the feature of brain electrical signals,decrypt potential information hidden in brain signals.An important aspect of EEG signal analysis is to understand EEG signals by extracting useful information.Using these extracted features,EEG signals can be classified or regressed according to the patterns and anomalies they reflect.Therefore,how to extract features more effectively and translate them into useful information has become a hot topic at home and abroad.At present,the feature extraction of EEG signals mainly use machine learning methods.This method does not need too much artificial participation.Only through training can we obtain effective features.This paper is based on machine learning methods to extract the feature.For the P300 signal,firstly,we propose a Semi-supervised Discriminant Analysis method to extract feature,and then,we propose a Semi-supervised Regularized Discriminant Analysis method,both methods improve the recognition accuracy under small sample conditions.For the EEG signals in anesthesia surgery,a model of EEG signal feature extraction and anesthesia depth monitoring based on long-term and short-term memory networks was designed.With the advantage of this model for timing signals,the feature is extracted and the depth of anesthesia is monitored.The main works and contributions of this article are mainly in the following three aspects:(1)A feature extraction method based on Semi-supervised Discriminant Analysis was proposed.This method is based on the original linear discriminant analysis(LDA),and optimize the objective function of LDA.Firstly,a relationship matrix is used to connect labeled data with unlabeled data,and then the relationship matrix is used as a penalty term to optimize the objective function to obtain a new objective function.The method is tested on the BCI Contest III data set,and the recognition accuracy is better than the traditional method.(2)A method of contraction Semi-supervised Regularized Discriminant Analysis method was proposed.This method is based on the Semi-supervised Discriminant Analysis(SDA),we use a shrinkage approach to optimize the total scatter matrix.The experiment proves that this method improves the recognition accuracy under small sample conditions.(3)A method of monitoring depth of anesthesia based on Long-Short Term Memory(LSTM)was designed.This method firstly preprocessed the EEG signals,and then extracted the features of EEG signals as the input of the LSTM network model,the multiple features was merged into one feature by this model,and the features was finally subjected to regression processing to obtain an anesthetic depth index.Experimental results show that this method has higher prediction accuracy.It has found the direction for the further development of monitoring depth of anesthesia.