Scholarlink

With the advancement of neuroscience and artificial intelligence, emotion recognition based on electroencephalogram (EEG) signals has attracted increasing attention in educational settings. To improve real-time monitoring of students’ emotional states in classroom environments, this study designs and compares three convolutional neural network architectures suitable for preliminary local feature extraction from EEG signals. Long Short-Term Memory (LSTM) networks are integrated for temporal modeling, and a multilayer perceptron is employed for valence binary classification. Experiments conducted on the DEAP dataset demonstrate that the EEGNet-LSTM model achieves an average classification accuracy of 77.9% and an F1 score of 72.9% across 32 subjects, outperforming ResNet-LSTM and Spatiotemporal Convolution-LSTM models in both feature extraction capability and model efficiency. This model is applicable to practical teaching scenarios and provides a technical foundation for EEG-based emotion recognition systems in education.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Lightweight CNN-LSTM Models for EEG-Based Emotion Recognition in Education

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Jing Wu,
  • Ningyuan Yu,
  • Wei Jin,
  • Qiming Zhao,
  • Jiaxing Luo,
  • Shijie Ling,
  • Wei Hu

摘要

With the advancement of neuroscience and artificial intelligence, emotion recognition based on electroencephalogram (EEG) signals has attracted increasing attention in educational settings. To improve real-time monitoring of students’ emotional states in classroom environments, this study designs and compares three convolutional neural network architectures suitable for preliminary local feature extraction from EEG signals. Long Short-Term Memory (LSTM) networks are integrated for temporal modeling, and a multilayer perceptron is employed for valence binary classification. Experiments conducted on the DEAP dataset demonstrate that the EEGNet-LSTM model achieves an average classification accuracy of 77.9% and an F1 score of 72.9% across 32 subjects, outperforming ResNet-LSTM and Spatiotemporal Convolution-LSTM models in both feature extraction capability and model efficiency. This model is applicable to practical teaching scenarios and provides a technical foundation for EEG-based emotion recognition systems in education.