EEG Signal Classification With Attention-Weighted Time-Frequency Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing motor imagery-brain computer interface (MI-BCI) systems face challenges in accurately classifying electroencephalogram (EEG) signals for motor imagery due to the complexity of integrating time-domain, frequency-domain, and spatial features, leading to suboptimal performance in predicting motor imagery types.
Innovation Solution
A method involving obtaining time-frequency feature maps from multiple electrode signals, performing feature extraction and spatial fusion, and applying an attention mechanism for weighting processing to enhance the classification accuracy of EEG signals, utilizing a convolutional neural network and attention weighted networks to integrate time, frequency, and spatial features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If separate model training andhyperparameter tuning are performed for each subject, then classification accuracy for individual subjects is improved, but training time and resource consumption increase significantly
Solution Approach 1:
The patent develops a universal EEG signal classification model that can be applied across multiple subjects without requiring separate training for each subject. The model uses domain adaptation technology to learn from a source domain (labeled data from one or more subjects) and apply it to a target domain (new subjects) with minimal or no additional training, thereby achieving multi-subject applicability and eliminating the need for individual model calibration
Solution Approach 2:
The patent employs domain adaptation to create a copied or adapted version of the source domain model that can be applied to the target domain. By learning the mapping relationship between source and target domains, the model generates a transferred classification model that replicates the source domain's performance on new subjects without requiring retraining, thus reducing time consumption while maintaining accuracy
2Reliability
If separate model training is performed for each subject, then individual classification performance is optimized, but device complexity and operational burden increase
Solution Approach 1:
The patent creates a universal classification model that serves multiple subjects simultaneously. The model incorporates domain adaptation mechanisms that allow it to maintain high classification performance across different subjects without requiring subject-specific training procedures, thereby simplifying the overall system operation and reducing calibration complexity while preserving reliability
Solution Approach 2:
The patent implements a self-adapting model that automatically adjusts to new subjects through domain adaptation techniques without requiring manual retraining or calibration. The model learns the domain mapping relationship and autonomously applies it to classify signals from new subjects, eliminating the need for complex manual calibration procedures while maintaining consistent performance
3Ease of manufacture
If traditional EEG classification methods are used, then implementation is straightforward, but generalization capability across different subjects is poor
Solution Approach 1:
The patent introduces a domain adapter as an intermediary component between the source domain model and the target domain application. This adapter learns the mapping relationship between different subjects' EEG signal characteristics and translates the source domain knowledge to the target domain, thereby enabling the model to generalize to new subjects while maintaining a relatively simple implementation structure
Solution Approach 2:
The patent employs parameter adaptation techniques that allow the model to adjust its internal parameters automatically when encountering new subjects. Through domain adaptation, the model modifies its feature representations and classification parameters to match the target domain characteristics, thereby achieving good generalization capability without fundamentally changing the model architecture or implementation approach
Data Source
AI summary
An electroencephalogram (EEG) signal classification method and apparatus, a device, a storage medium, and a program product are provided, and relate to the field of signal processing technologies. The method includes: obtaining a first EEG signal; obtaining time-frequency feature maps of at least two electrode signals in the first EEG signal; performing feature extraction based on the time-frequency feature maps of the at least two electrode signals to obtain a first extracted feature map; performing weighting processing based on an attention mechanism on the first extracted feature map to obtain an attention feature map; and obtaining a motor imagery type of the first EEG signal based on the attention feature map.