AMFTCNet: A multi-level attention-based multi-scale fusion temporal convolutional network for decoding MI-EEG signals

Abstract

Currently, Brain-Computer Interface (BCI) technology based on Electroencephalography (EEG) and Motor Imagery (MI) is widely applied in fields such as neural rehabilitation, assistive communication, and virtual reality. However, MI-EEG signals are susceptible to various factors that affect decoding accuracy and generalization capability. Effectively utilizing training data and integrating signal features remain major challenges in MI-EEG decoding algorithms. To address these issues, this paper proposes a novel AMFTCNet model. The model improves decoding results through multi-scale feature learning and dynamic integration of features from different scales. The AMFTCNet model first extracts multi-scale feature representations from the raw EEG signals using Convolutional Blocks (CV) and Multi-Scale Branch Structures (MSB). It then extracts high-dimensional features from single scales through Parallel Attention Temporal Convolution Blocks (PAT). Additionally, this paper introduces a new attention block, the PSCA block, which dynamically weights and combines high-dimensional features from different scales to integrate signal features and enhance decoding performance. Experimental results demonstrate the superior performance of the AMFTCNet model across multiple datasets. The model achieves accuracies of 87.77%, 88.26%, and 95.62% on the BCI Competition IV-2a, BCI Competition IV-2b, and High Gamma datasets, respectively. These results provide valuable insights for exploring how to effectively fuse multiple types of feature information and how to utilize attention mechanisms more efficiently to improve decoding accuracy.