| With the rapid development of brain computer interface technology,emotion recognition has become a research hotspot in this field.EEG acquired by wearable devices can effectively monitor human emotions.In recent years,traditional machine learning methods and deep learning methods have achieved good classification results in emotion classification tasks based on EEG signals.However,in practice,they are difficult to be applied in wearable devices due to the large number of channels.In some researches based on channel selection,good classification results can often be obtained in a small number of channels through the selection of channels by algorithms.However,the dataset used in these studies are not uniform,which makes the results lack of extensive comparison;Moreover,in the case of different subjects,different time periods and different stimulation materials,the number and location of the optimal channel are uncertain,which hinders the application of the algorithm in wearable devices.In this thesis,a method of transforming the original channel signal into the basic signal under the condition of fixed 4 channels is proposed to preserve the spatial domain information of brain region as much as possible.At the same time,a parallel processing method based on discrete wavelet transform,inherent time scale decomposition,variational mode decomposition and phase space reconstruction is proposed to obtain more modes of the signal.On this basis,the differential entropy of the signal is extracted as the feature,and the extracted feature is smoothed in time sequence by using the linear dynamic system.Based on the smoothed artificial features,this thesis uses support vector machine and convolutional neural network model designed in this thesis to carry out three emotion classification experiments on SEED dataset and the dataset collected in this thesis.The experimental results show that the average accuracy of this method can reach the classification accuracy of more channels(62 channels of SEED dataset and 14 channels of the dataset collected in this thesis).In addition,based on 9vs6 protocol,the classification accuracy of SEED dataset in this thesis is about 5% lower than the current research results under multi-channel,which shows good classification performance and development potential.Finally,based on the method proposed in this thesis,an emotion recognition system based on EEG signal is developed.Because this method has a good classification accuracy in the case of fixed 4 channels,it is expected to be applied in wearable devices with weak computing power.At the same time,with the help of emotion recognition system,this method can be popularized in a wider range. |
