Research On SSVEP Recognition Based On Compact Neural Network And Its Applications

Steady state visually evoked potential(SSVEP)is a common brain-computer interface(BCI)paradigm,which is an EEG signal generated in the posterior occipital region of the brain after the subject receives a stimulus with a fixed flicker frequency that is correlated with the frequency of the stimulus target and its harmonic frequency.SSVEP has received a lot of attention from researchers because of its convenient acquisition and high information transmission rate.In order to improve the performance of neural networks to decode SSVEP signals,the main work of this paper are as follows:1)A multi-task model based on compact neural network is proposed: While traditional machine learning algorithms have achieved good performance,to further extract more complex nonlinear features in SSVEP signals,researchers have begun to turn their attention to neural network models.However,due to the small size of the SSVEP dataset,it is difficult for the existing model to learn some effective features.In order to make the neural network more effective in extracting the features of SSVEP signals,this paper proposes a multi-task model based on compact neural network: MTEEGNet.The model enhances the feature extraction capability of the middle layer of the model by adding a classification auxiliary task and adding a correlation-constrained auxiliary task to constrain the features extracted from the feature layer between the classification auxiliary task and the main task to be as similar as possible,and evaluates the performance of MT-EEGNet against mainstream methods in both cross-subject and subject-specifc conditions on both public dataset and experimental datasets.The results show that the MT-EEGNet model proposed in this paper effectively enhances the feature extraction capability of the middle layer of the model,which in turn improves the overall performance of the model.2)A compact neural network-based cross-stimulus target fusion model is proposed.To further extract harmonic information and cross-stimulus target information,this paper proposes a compact neural network-based cross-stimulus target fusion model: FBEEGNet.The model gives a certain probability weight to non-gazing targets by modifying the output labels,and is trained in stages to ensure that the network can effectively extract multi-band information while preventing model overfitting.To evaluate the performance of the model,FB-EEGNet is compared with some mainstream methods on both public dataset and experimental dataset under cross-subject and subjectspecific conditions in this paper.To evaluate the performance of the model,FB-EEGNet is compared with some mainstream methods on both public dataset and experimental dataset under cross-subject and subjcet-specific conditions in this paper.The results demonstrate that our method effectively extracts information from multi-band and nongazing target stimuli.3)A SSVEP-based hand exoskeleton control system was designed and implemented.A transfer learning approach enables this online exoskeleton control system to be used by collecting only a limited set of user data.The final performance of the system is verified by online hinted and unhinted experiments,demonstrating that the system only needs to use a limited set of user data to train the model to obtain a performance very close to that of the subject-specific system.