A dual-branch motor imagery deep learning classification method

By combining Gammatone filters and multi-scale spatiotemporal convolutional networks, a Gammatone-Shuffle Attention Network (GSANet) is constructed, which solves the problem of insufficient feature extraction in MI-EEG signal classification of existing models and achieves high-precision and robust MI-EEG signal decoding.

CN122241368APending Publication Date: 2026-06-19NANCHANG UNIV
0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
NANCHANG UNIV
Filing Date
2026-03-23
Publication Date
2026-06-19

AI Technical Summary

⚠Technical Problem

Existing deep learning models struggle to efficiently extract features from different frequency domains in decoding motor imagery EEG signals and are prone to losing complex spatiotemporal structural features, resulting in insufficient accuracy and robustness in MI-EEG signal classification.

⚗Method used

By combining Gammatone filters with multi-scale spatiotemporal convolutional networks, a dual-branch Inception spatiotemporal convolutional network is constructed. Combined with a sliding window attention mechanism and lightweight convolution, frequency domain feature extraction and spatiotemporal feature fusion are achieved, and a Gammatone-Shuffle Attention Network (GSANet) is constructed for classification.

🎯Benefits of technology

It significantly improves the decoding accuracy and robustness of MI-EEG signals, enhances the robustness and generalization ability of the model, can efficiently capture the spatiotemporal frequency information of MI-EEG signals, and improves the feature representation capability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 1
    Figure 1
  • Figure 2
    Figure 2
  • Figure 3
    Figure 3
Patent Text Reader

Abstract

This invention discloses a two-branch deep learning classification method for motor imagery, belonging to the fields of neuroscience and computer technology. S1: Preprocess the continuous EEG signals of the subject to obtain standardized EEG signals; S2: Construct a Gammatone-filtered convolutional layer based on the preprocessed EEG signals; S3: Construct a two-branch Inception spatiotemporal convolutional network based on the frequency domain feature signals output from S2; S4: Construct a sliding window attention mechanism module based on the fused feature maps from S3, process the feature maps of each window, and connect the feature maps along the channel dimension to form integrated features; S5: Construct and train a classification model based on the integrated features from S4, mapping the features to class probabilities to achieve classification of motor imagery EEG signals. Using this method, the classification performance of motor imagery tasks can be significantly improved, greatly promoting the practical application of brain-computer interfaces and enhancing the robustness and generalization ability of the model.
Need to check novelty before this filing date? Find Prior Art