EEG Time-Frequency Feature Optimization Algorithm Based on Continuous Wavelet Transform and CNN Model

Brain-Controlled technology enables direct interaction between the brain and peripheral devices by decoding electroencephalographic (EEG) signals from subjects. Among these, brain-controlled method based on hand movements have attracted widespread attention from scholars due to their fast response speed. However, traditional methods for decoding EEG signals associated with hand movements are ineffective in feature extraction, making it difficult to accurately represent subject intentions in weak EEG signals. To address this issue, this paper proposes a Convolutional Neural Network based on Continuous Wavelet Transform (CWT-CNN) method that combines reinforcement learning and transfer learning for feature extraction and classification of EEG signals from hand movements. This method first utilizes continuous wavelet transform to extract time-frequency features, then uses reinforcement learning neural architecture search (RL-NAS) to generate the NAS-CNN model structure with the best classification performance as a pre-trained model, and finally fine-tuning its parameters through transfer learning to construct the CWT-CNN model, thereby achieving efficient classification of six-class hand movements EEG signals in two-dimensional time-frequency image (2D-TFI). To verify the effectiveness of proposed method, the study recruited a total of 8 participants for offline experiments. The averaged recognition accuracy is 71.63 ± 2.98% and the highest one is up to 74.65%. The experimental results demonstrate that the CWT-CNN method proposed in this study effectively improves the decoding accuracy of EEG signals from hand movements, fully proving the application value of the CWT-CNN method, in enhancing the recovery of activities of daily living in stroke patients.