The Study Of Task-Free Brainprint Recognition

Personal identification plays an important role in the information society.However,the traditional methods cannot guarantee the security completely.Brainprint refers to the brain produced by a unique,with a collection and persistence,and can be used for identification,verification of EEG signal characteristics.It has obvious advantages such as strong concealment,non-stealing,and can only be detected in vivo.It has a broad application prospect in fields with high requirements on confidentiality and security.At present,the brainprint recognition system required the subjects to carry out under specific tasks.Usually,external stimulation or specific thinking and imagination were required for the identification.This has caused a great obstacle to the popularization of the brainprint recognition,so the study of task-free brainprint recognition is particularly important.Task free refers to the fact that does not depend on the task type,that is,a method of brainprint recognition can target the EEG signals in various cognitive task states.In this thesis,three different methods are used to study task-free brainprint recognition:(1)Feature extraction method based on phase synchronization and degree of node was studied.In this thesis,the relationship between phase synchronization and brain physiological characteristics was used to analyze the phase synchronization characteristics between two electrodes.EEG signals were segmented in a certain time length,and then the average phase locking value within the time period was calculated.We took the phase locking value as the weight between nodes of the brain functional network,used the generated weight matrix to create a weighted undirected network.The degree of node was used to extract features for personal identification.(2)A task-independent background EEG identification method based on low-rank matrix decomposition(LRMD)was studied.We assumed that task-related EEG data could be separated into two parts,the background EEG(BEEG),the residue EEG(REEG).The BEEG contained intrinsic features unique to a subject,representing low-rank characteristic.The REEG contained random noise and task-evoked EEG(TEEG)which represented a random signal generating from the corresponding cortical neurons activated by a task.To begin with,the original EEG data was transformed into EEG-spectrogram via short time Fourier transform,GoDec+combined with fast rational quadratic kernel was used in the BEEG extraction.The ensemble subspace of all subjects was constructed according to the rank of BEEG data matrix.Finally,the best reconstruction coefficient was calculated based on maximum correntropy criterion,and the test samples were classified.(3)The recognition method based on convolutional neural network model was studied.The model structure consisted of 8 layers:input layer,3 convolution layers,1 pooling layer,1 local response normalization layer,1 fully connected layer and output layer.A multilayer neural network model was constructed by convolutional neural network for unsteady time series multichannel EEG signal.This method could recognize EEG signals under different task states and estimate the identity of each subject accurately,efficiently and objectively.Experiments were carried out on four data sets,including multi-task EEG dataset,SJTU emotion EEG dataset,motor imagery-based BCI2008 dataset and event related potential-based P300 EEG dataset.The above methods achieved good recognition results.This showed that our methods could not only adapt to the changes of the task,but also to some extent resisted the interference of emotions on identification.