A Lightweight 1D-CNN for Emotion Recognition Based on EEG Multi Feature Fusion

Emotion detection based on EEG is more objective and reliable compared to human actions and language. Recent advancements in deep learning technology have enabled the successful application of convolutional neural network models in EEG-based emotion recognition, yielding favorable outcomes. Nevertheless, high-performance convolutional neural network models are often intricate in structure and necessitate substantial computational and storage resources, posing a significant challenge for resource-constrained embedded hardware to accommodate. This article proposes a relatively simple and efficient EEG signal preprocessing method to address this issue. Multiple feature sequences are extracted from the raw signals of EEG channels and fused to form feature sequence samples. We designed a lightweight 1D-CNN neural network model and conducted binary and quaternary classification task experiments on the DEAP dataset. We achieved accuracies of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$97.44 \%, 97.57 \%$</tex>, and 96.14 % for Valence, Arousal, and Arousal-Valence classification tasks, respectively, verifying the effectiveness of the EEG signal multi feature fusion preprocessing method on the 1DCNN neural network model. Finally, using quantization after training, the model was compressed to 0.562 MB. The experiment showed that our dataset preprocessing method ensures the effectiveness of feature sequences while consuming less computational resources. At the same time, our proposed model is lighter and has lower computational complexity compared to most research structures, making it more hardware friendly.