Adaptive Multi-Scale CNN–Transformer with Hybrid Attention Gating for Reliable EEG Emotion Recognition
Abstract
Electroencephalography(EEG) based emotion recognition has developed as a decisive field of inquiry into emotional analysis and brain-computer interface (BCI) systems because of its capacity to provide accurate and consistent data of human emotional states. However, the varying characteristics of EEG signals, involved time series, and cross-channel dependencies present substantial challenges in the accurate classification of emotions. Conventional convolutional neural networks mainly capture local features, whereas existing transformer-based models frequently demonstrate redundant feature representations and confined dynamic variable reduction. To solve these problems, this study proposes a hybrid multiscale dynamic CNN-transformer framework integrated with a Hybrid Attention Gate (HAG) and Temporal Convolutional Network (TCN) for EEG-based emotion recognition. The proposed Multi-scale Dynamic CNN identifies distinctive temporal and spatial features using multiple convolution kernel sizes to capture EEG characteristics among various frequency bands. A Transformer encoder is consequently used to model Long-term connections through multi-head self-attention, while the proposed HAG adaptively strengthens informative features and reduces redundant representations using an attention-guided gating mechanism. Finally, a TCN further models the sequential temporal dependency before the softmax classification predicts the emotional category. The proposed architecture was examined on the accessible DEAP dataset using a five-fold cross-validation. The findings establish that the proposed model accomplish a high accuracy of 99.12%, exceeding conventional CNN-, Transformer, and hybrid deep learning methods.