Research On EEG Signal Recognition Algorithms Based On Wavelet Transform And Deep Learning

Electroencephalogram(EEG)collects data on changes in brain potential signals.EEG is mainly used for diagnosis and monitoring of neurological disorders such as epilepsy and Alzheimer’s disease.It is of great importance for human to understand the working mode of the brain nervous system and to realize human-computer interaction.However,due to the nonlinear and non-smooth characteristics of EEG signals,the fluctuation characteristics of EEG signals are difficult to identify and thus affect the diagnosis.To solve this problem,this paper combines feature extraction theory and convolutional neural network algorithm based on adjustable Qfactor wavelet transform theory.Multiple learning algorithms are constructed for different situations to improve the EEG signal recognition accuracy and better solve the above problem.The main work of this paper is as follows:(1)To solve the problem that nonlinear and non-smooth features of EEG signals are difficultly and inadequately extracted,we propose an adaptive tunable Q-factor wavelet transform with multi-feature(Ad-TQWT-MF)in this paper.The EEG signal is decomposed by defining the energy-Shannon entropy ratio as the evaluation of the decomposition,and selecting the appropriate Q-factor to decompose the EEG signal.The relationship between the decomposed subband information and the original EEG signal is analyzed to extract the EEG signal time domain,frequency domain and nonlinear multidimensional features.Finally,adaptive feature selection is used to filter features and construct a feature subspace that helps EEG signal classification study.Experiments are conducted on the imaginary EEG signals of the motor movements of the BCI-III competition.Compared with the traditional wavelet transform method,the Ad-TQWT-MF algorithm extracts the different features of the EEG signals more fully,which is more suitable for the EEG signal recognition algorithm and has a higher accuracy rate.(2)The advantages of the Ad-TQWT-MF algorithm in EEG signal feature extraction are further investigated.It is found that if the link between the Q-factor selection criteria and wavelet sub-band fluctuation features is increased,the signal recognition accuracy will be further improved.Therefore,a Revised tunable Q-factor wavelet transform(RTQWT)algorithm is proposed based on the Ad-TQWT-MF algorithm.The RTQWT uses the weighted normalized entropy of the subbands obtained from wavelet decomposition to modify the Qfactor.The energy distribution of the subbands after wavelet decomposition is calculated,and the information on the time-frequency characteristics of the EEG signal in each wavelet subspace is measured.Then,the frequency domain distribution of the EEG signal subbands is combined with the characteristics of the frequency domain distribution of the EEG signal to construct the unique and common feature subspaces of EEG signals.The final machine learning algorithm is used to complete the classification and identification of EEG signals.The RTQWT was experimented on Bonn epilepsy EEG signals.The experimental results show that the RTQWT can effectively handle the fluctuations of EEG signals.The special fluctuation features contained in different seizure states in the epilepsy EEG signal can be accurately extracted,which in turn can rapidly improve the EEG signal recognition accuracy.(3)Although the wavelet transform based on adjustable Q-factor can better extract various signal features of EEG signals,traditional machine learning algorithms cannot tap the depth features in EEG signals,which leads to weak generalization ability of classification algorithms.In particular,when facing deep learning(e.g.convolutional neural network)algorithms,there are problems of many model parameters and slow training speed.We propose a Multihierarchical attentional convolution neural network(MHA-CNN)algorithm incorporating RTQWT in order to solve the above problems and introduce an attentional mechanism based on the convolutional neural network model.First,the RTQWT algorithm decomposes the EEG signal and extracts the fluctuating features contained in the signal to form a pre-trained feature set.This feature set is then placed into a modified CNN structure embedded with multi-layer attentional reinforcement for training.The special fluctuating features in the EEG feature set are reinforced by the attention mechanism at different levels to accelerate the CNN network training process.The convergence speed of the overall MHA-CNN algorithm for EEG signal data training is improved.The results of related experiments show that MHA-CNN has fewer iterations and faster and more accurate recognition of EEG signals.The classification accuracy of EEG datasets reaches 96.7% to 99.2%.