Research On Emotion Recognition Method Based On Recurrent Fuzzy Neural Network And EEG Signals

Emotion not only reflects the current physical and psychological state of people,but also plays an important role in human cognition,communication and decision-making ability.Emotion is a necessary condition for human social activities.Emotion recognition can employ computers to automatically recognize,understand and reflect human emotions,thereby assisting in all aspects of real life,such as human-computer interaction,medical care,education,game development,driving safety and so on.Considering that EEG signals can provide powerful objectivity and high classification accuracy for emotion recognition,as well as the rapid development of brain-computer interfaces,the research of EEG emotion recognition is a major topic given to us by the times.Through research and found that the existing EEG emotion recognition research has achieved a good emotional recognition effect,but can still be further improved in the following areas.(1)The existing research ignores the fact that EEG signals labeled as “positive” or “negative”emotions may contain neutral emotion signals,which indicates that there are data with mismatched labels in EEG emotion recognition datasets.This affects the accurate training of emotion recognition model.(2)The network structure of the existing studies lacks the ability to take full account of the emotional EEG characteristics.This study defines the correlation of EEG signals from multiple electrode channels,the correlation of EEG signals from the same electrode channel in different frequency bands,the ambiguity and the time continuity of EEG signals as the characteristics of emotional EEG signals.The ambiguity means that an emotional feature does not correspond to a specific value,but corresponds to a range of values.The time continuity refers to the continuity and correlation of emotions in time.In fact,fuzzy logic can effectively deal with ambiguity,convolutional neural networks can effectively extract emotional spatial information,and cyclic neural networks can effectively extract emotional temporal information.However,few models combine fuzzy inference with convolutional neural networks and recurrent neural networks for emotion recognition.To address the imperfections,this paper proposes a semi-supervised learning convolution recurrent fuzzy neural network(SCFRNN).First,the semi-supervised learning mechanism is employed to filter out the mislabeled EEG signals in the EEG dataset.Thereby optimizing the EEG dataset.Second,vectorization convolution is adopted to extract the relationship between the EEG channels and the frequency bands.It pays attention to the relationship between EEG signals in multiple channels and multiple frequency bands,which extracts information in line with the structure of the brain.Third,the fuzzy reasoning is combined with the recurrent neural network to focus on the ambiguity and time continuity of EEG signals,which helps to extract EEG signals' effective temporal information for better understanding emotion.Because of the rational combination of the three schemes,SCFRNN can exhibit competitive accuracy and robustness on SEED(SJTU Emotion EEG Dataset,SEED).