Research On Cross-domain Decoding Method Of Spontaneous EEG Signals

Brain-computer interface enables direct communication from the central nervous system to machines by establishing a pathway between humans and computers.As the basic technology in the field of human-computer interaction in the future,it is gradually showing broad application prospects.Brain-computer interface can be divided into spontaneous and evoked types.Spontaneous brain-computer interface is highly anticipated as the most ideal form of brain-computer interface.At present,spontaneous brain-computer interfaces have produced some exciting innovative products in markets such as medical rehabilitation,education,and entertainment games.However,as the core technology of spontaneous braincomputer interface,the spontaneous EEG signal decoding method still has some problems that need to be overcome.The existing spontaneous brain-computer interface requires a long time to collect the user’s EEG data for training and calibration of the decoding model before use,which undoubtedly brings a poor user experience.Especially in the fields of medical rehabilitation and education,users often do not have the energy to collect training data,which greatly limits the application of brain-computer interface technology and products in these fields.The ideal way to solve the above problems is to train and calibrate new subject models using cross-domain data from existing subjects.The cross-domain decoding of spontaneous EEG can be divided into cross-trial,cross-subject and cross-device.Through in-depth study of the physiological basis and existing methods of cross-domain problems,this article proposes new solutions to cross-domain problems at three levels.Its main research works and innovations are summarized as follows:(1)Spontaneous EEG cross-trial decoding method with function predefined convolution.Spontaneous EEG cross-trial decoding is the most basic cross-domain task.The difficulty is mainly reflected in the changes in the EEG signal between different trials due to external interference,subject state,and scalp contact conductivity.This difference generally does not cause significant changes in the spatiotemporal distribution of signal characteristics,but mainly manifests itself as changes in signal intensity and signal-to-noise ratio.Since the mapping space of existing neural network methods includes not only the feature space of task-related rhythm information,but also part of the feature space of physiological signals and noise.It is very easy to produce overfitting between trials,thus affecting cross-trial classification accuracy.In order to solve the above problems,this work innovatively proposes a spontaneous EEG cross-trial decoding method with function predefined convolution.This method combines the advantages of the computational efficiency of traditional methods and the learning ability of neural networks,designs a novel function predefined convolution layer,and builds a function predefined convolutional neural network based on it to decode spontaneous EEG signals.Experimental results show that compared with the state-of-the-art comparative methods,the cross-trial decoding performance of the proposed method on three spontaneous EEG data sets is improved by 2.09%,3.08% and3.41% respectively.At the same time,the number of trainable parameters is reduced by99.96%,99.99% and 92.17% respectively.(2)Domain-adaptive spontaneous EEG cross-subject decoding method.Cross-subject decoding of spontaneous EEG is the most valuable cross-domain task.The difficulty is mainly reflected in the cross-subject differences caused by differences in brain structure,imagery ways and data recording between different subjects.The above-mentioned data differences mainly focus on spatial differences reflecting the location of brain activity and feature distribution differences reflecting the way of imagery between subjects.This work proposes a novel domain-adaptive spontaneous EEG cross-subject decoding method based on an in-depth analysis of the above difficulties.This method proposes a new feature extractor,spatial discriminator and conditional discriminator on the network structure.The feature extractor uses a spatio-temporal convolutional network design,which has a more reasonable structure and fewer parameters,while reducing the risk of over-fitting across subjects.the spatial discriminator and conditional discriminator iterate the feature extractor during the training process,thereby helping the feature extractor learn common feature representations across subjects.In the experimental part,the performance of the proposed method was tested on two public spontaneous EEG datasets using two transfer protocols:one-to-one transfer and leave-one-out transfer.For the one-to-one transfer protocol,the classification accuracy of this method is improved by 5.70% and 12.43% over the baseline method.For the leave-one-out transfer protocol,the accuracy improvements are 8.60% and 15.84% respectively.Experimental results show that the cross-subject spontaneous EEG decoding performance of this method is significantly improved compared to existing methods.(3)Multi-source domain sample reweighted spontaneous EEG cross-device decoding method.Cross-device decoding of spontaneous EEG is the most promising cross-domain task.The difficulty lies in the significant differences in EEG data of different subjects collected by different devices.And when multiple subjects across devices serve as multisource domain training sets,the feature distribution differences and classification interfaces of multi-source domain data will be highly mixed and scattered.How to automatically select appropriate training samples from multi-source domain data based on the data in the target domain has become the key to solving cross-device decoding.This work innovatively proposes the idea of multi-source domain reweighted training based on in-depth exploration of the characteristics of multi-source domain data.Different from the traditional idea of domain adaptation to improve cross-domain decoding accuracy by confusing feature distribution,this work considers that target domain features can be represented as a weighted combination of multi-source domain features.Under the guidance of this idea,a multi-source domain reweighted training strategy and corresponding neural network structure are proposed.A novel sample reweighting classifier is designed in the network,which rereweights the training weights for multi-source domain samples by measuring the distance between multi-source domain samples and the feature center of the target domain.Thereby training the samples that are closer to the target domain feature distribution with high training weights.At the same time,a sample reweighted discriminator is used to reweight the training weights of multiple source domain samples based on classification confidence,thereby achieving high training weights for source domain samples that are close to the target domain category interface.This work validates the method performance by conducting cross-device decoding experiments on three publicly available spontaneous EEG datasets.In three crossdevice transfer experimental tasks,the proposed method achieved decoding accuracy of 6.88%,5.90% and 3.49% better than the baseline respectively.The experimental results confirm the effectiveness of the sample reweighting training strategy and the corresponding network structure for spontaneous EEG data cross-device decoding.