A CNN Approach in Building EEG-Based Emotion Recognition System for AIoT Applications

In practice, when deploying AIoT applications for emotion recognition based on EEG signals, the devices not only need to be mobile and wireless but often have a small number of electrodes to ensure portability, compactness, and energy efficiency. Additionally, the classification blocks using deep learning networks also need to be adjusted to achieve architectures with fewer parameters while maintaining effective recognition capabilities. Therefore, improvements to the deep learning architecture are necessary to meet these criteria. In this paper, we propose an architecture named EEG_MCIS, which is a 1D-CNN enhanced with Inception and Squeeze and Excitation blocks to suit the EEG data collected from devices with a small number of electrodes. The Inception block divides the 1D-CNN into two branches. The first branch with small kernel size aims to extract local features from the data, and the other with a larger kernel size to enable the model to perceive features over a wider range and capture relationships between more distant data points along the time axis. Meanwhile, the Squeeze and Excitation block leverages information about the importance of each channel, optimizing the classification process of complex EEG data. Experimental results aimed to compare with recent publications on standard datasets (DEAP, AMIGOS, DREAMER) show that the proposed EEG_MCIS architecture has fewer parameters and achieves better recognition efficiency, with an accuracy rate of up to 99% even on EEG devices with only 5 electrodes.