Separable Convolution Fusion Representation of Four-dimensional Features and Learner Cognitive States Recognition Based on EEG Signals Study

The development of higher-order thinking skills, a key objective in education and teaching, lies at the heart of learners’ competence development. Accurately identifying learners’ cognitive states is crucial for educators to optimize teaching strategies and develop personalized learning. Research has shown the efficacy of using features derived from electroencephalogram (EEG) signals to detect cognitive states. However, effectively fusing these features is a challenge for current research. In this study, we examine a cognitive state recognition method for learners by integrating time-domain, spatial-domain, and frequency-domain features of EEG signals. Initially, learning activity experiments were designed to elicit various cognitive states from learners during the learning process, with EEG signal data collected from participants engaged in these activities. Subsequently, we investigated the fusion of temporal, spatial, and frequency domain features of EEG signals, proposing a four-dimensional (4D) feature representation and recognition model based on depthwise separable convolutional fusion, named MF-DSC-BiLSTM. The model was tested on our self-constructed Learner Cognitive State Dataset (LCSD) and the public sentiment dataset SEED, achieving recognition accuracies of 94.24% and 96.80%, respectively. These experimental results indicate that the MF-DSC-BiLSTM model is capable of effectively identifying learners’ cognitive states during the learning process.