Research On Event-related Potential Classification Method Based On Convolutional Neural Network

Brain-computer interface(BCI)is a direct connection created between the human or animal brain and an external device through which information can be exchanged between the brain and the external device.This form of interaction can be used as additional language and body movements as a means of information exchange,and it has great potential for helping people with motor or sensory disabilities to regain information exchange functions.Event-related potential P300 is a special type of electroencephalogram(EEG)signal applied in brain-computer interfaces,which is generated when the nervous system is stimulated in a specific way and can be detectable in the corresponding part of the brain.Since the event-related potential P300 itself is relatively distinctive,it is widely used in brain-computer interface systems.Character spelling systems are an important application of BCI systems based on the event-related potential P300,which allows patients with disabilities to communicate with others via BCI.As a non-invasive brain-computer interface system,Character spelling systems is a popular brain-computer interface paradigm with the advantages of low price and high safety.Due to the non-invasive nature of its signal acquisition and the characteristics of event-related potentials themselves,the acquired event-related potential P300 signals have low signal-to-noise ratio,unbalanced sample distribution,and variable latencies and peaks,making it difficult to improve the resolution of braincomputer interface systems.Deep convolutional neural networks have been used for the classification of event-related potential P300,but the current classification methods still need to be improved in terms of their ability to classify the signals.To address these issues,the work in this thesis is described below.(1)To address the problem of unbalanced sample distribution in event-related potential datasets,which can affect the training of classification algorithms,this thesis proposes an integrated data enhancement method to balance the datasets.This thesis integrates existing data enhancement methods to improve the effectiveness of data enhancement methods and minimize the impact of unbalanced event-related potential datasets on classification methods.(2)To address the problem that the current deep learning network is not strong enough to classify event-related potential signals,this thesis proposes a deep learning network called CIE-EEGNet,which enhances the fitting ability of the network by strengthening the ability of combining features between different channels and increasing the representation of low-level features in the network,so as to improve the classification ability of the deep network to improve the convolutional neural networks to improve the classification accuracy of P300 signals.(3)To address the characteristics of latency and peak instability of event-related potential signals in the time domain,this thesis combines parallel convolutional layers with attention mechanism to further improve the generalization ability of the network for classification of event-related potential P300.The EEG signals are first convolved by parallel convolutional layers with different convolutional kernel sizes,and then the acquired features are connected.Since convolutional kernels of different sizes have different receptive fields,more features are obtained compared to one convolutional layer.The fused features are then reinforced using an attention mechanism to suppress the noise while enhancing the representation of useful features.The expressiveness of the network with the addition of multiscale feature fusion and the attention mechanism is further improved and the classification ability.