Emotional recognition of EEG signals utilizing residual structure fusion in bi-directional LSTM

Emotion recognition using electroencephalogram (EEG) signals had attracted significant research attention. This paper introduced a new approach, Multi-scale-res BiLSTM (MRBiL), to enhance EEG emotion recognition. Firstly, differential entropy features were extracted from EEG data during both resting and active states. The disparity between these two states was then calculated to generate a feature matrix, which was subsequently input into a multi-scale convolution block. The high-dimensional feature matrix extracted from the convolution block was mapped using a residual block. The feature signal sequence was then processed by a bidirectional long-term short-term memory network. Finally, the emotion recognition result was obtained through the classification layer. The algorithm was validated in the DEAP dataset, resulting in average accuracies of 0.9788 for binary classification of validity and arousal. Furthermore, the algorithm attained an average accuracy of 0.9685 for quadruple classification. Additionally, ablation experiments were conducted in this study to affirm the effectiveness of each deep learning component within MRBiL. The experimental results demonstrated that the novel neural network model proposed in this paper had outperformed currently available methods in emotion recognition tasks.