Classification study of motion intention signals based on multilevel feature approach

Brain-computer interface technology facilitates information exchange and communication between the human brain and external devices. By decoding the brain’s intentions through capturing its electrical signals (e.g., electroencephalogram (EEG)), it translates these intentions into instructions for controlling external devices. This technology holds promising applications in medical care and humancomputer interaction. In order to improve the accuracy of classification, many studies have solved the problem of long dependence of sequential signals by the mechanism of selfattention. This paper proposes a multi-level feature-based classification model for motion intent signals to address the issue of a single level of extracted features and poor performance, including a low correct classification rate. The model employs a structure of multi-level feature fusion and the self-attention mechanism, enabling it to capture a broader range of temporal and spatial features of sequence signals. The model demonstrated superior performance to the traditional CNNbased approach in both training and testing, achieving an accuracy of 83.10% on the BCI Competition IV-2a dataset.