EEG-based emotion recognition is one of the key technologies to improving the interaction between computers and the human brain, which can be applied in many industries such as healthcare, entertainment, and psychology. Compared to traditional emotion recognition features (facial, auditory), electroencephalogram (EEG) is considered to be more accurate and more reliable. Thus, in past few years, EEG-based subject-dependent emotion recognition has been intensively investigated using machine learning models such as SVM and KNN. However, many experimental results show that deep learning methods tend to be superior when applying to subject-independent emotion classification. In this thesis, we study a novel deep learning model architecture that utilizes autoencoder model structure to decompose original EEG data into several key signal components and power spectral density (PSD) is extracted, then LSTM recurrent neural network is used to capture the temporal relationship of PSD feature sequence. 66.95% and 70.00% accuracy of positive and negative emotion classification can be achieved in valence and arousal dimensions, respectively. A lot of comparison experiments have been done to try to find the optimal model structure and hyperparameters. Furthermore, we utilized the open-source Python package MNE to help us better understand, visualize, and analyze human EEG data.
