Research on Epilepsy Detection and Recognition Based on the Combination of Time Frequency Transform and Deep Learning Model

To improve the detection performance of epileptic electroencephalogram (EEG) signals and address their non-stationary characteristics, this paper compares the combined effects of continuous wavelet transform (CWT), short-time Fourier transform (STFT), along with three neural network models—EEGNet, AlexNet, and Shallow ConvNet—and incorporates innovative designs. Specifically, Focal Loss, dynamic data augmentation, and an early stopping mechanism are introduced at the training stage to enhance the model robustness. Additionally, EEGNet is optimized by integrating an SE (Squeeze-and-Excitation) attention module, improving depthwise separable convolution (where a (3,16) kernel is used in the first layer), and dynamically adapting dimensions to reduce errors. For Shallow ConvNet, improvements are made by adopting layered convolution to extract “time-frequency” features and average pooling to adapt to long data blocks. The results show that the recall rate of the CWT+Shallow ConvNet combination reaches 100% with an accuracy of 99.14%, while the accuracy of the CWT+EEGNet combination achieves 100%. These findings verify the effectiveness of combining precise time-frequency features with optimized models, providing support for clinical practice.