Eeg Signal Classification Combined With Spatial-temporal Analysis And Deep Learning

Brain-computer interface(BCI)is an important research field in human-computer interaction.It can reflect various functions of the human body through electroencephalogram(EEG)signals,including fatigue,emotion,healthy state,and so on.The study of EEG signals can not only promote the deeper and broader applications of human-computer interaction,but also help to understand the information processing mechanisms in human brain cognitive processes.EEG signals have a high temporary resolution,and contains rich information of inter-electrode correlations,which is the spatial-temporal information in EEG signals.Deep learning techniques have been proven useful in multiple image analysis fields with their strong abilities to learn vital representations from data.However,it has a very limited exploration in the field of EEG signal analysis.Based on multi-channel EEG signals,this paper aims to develop novel spatial-temporal-features-based deep learning models to extract stable taskrelated representations.The main works are as follows:1)A spatial-temporal convolutional neural network is developed to recognize EEG signals of the subjects under different driving fatigue conditions.We design the fatigue driving experiment to collect all the EEG signals of eight subjects from alert state to fatigue state during the simulated driving process,and develop a spatial-temporal convolutional neural network.Its main ideas can be summarized as follows: First,core blocks are introduced to extract the information across the temporal dimension from EEG signals;then dense layers are employed to fuse spatial features among the electrodes.Results show that this method combines the spatial-temporal information of EEG signals,and achieves an average accuracy of 97.37% and a standard deviation of3.30% in the fatigue dichotomy task.2)A channel-fused dense convolutional network is developed to recognize different emotion states of the subjects in the emotion EEG classification tasks.Here,we use two commonly used public emotion EEG datasets,SEED and DEAP,to build the channel-fused dense convolutional network.Its main ideas can be summarized as follows: First,1D convolution is used to receive weighted combinations of contextual features along the temporal dimension of EEG signals;then,1D dense structures are introduced to capture electrode correlations along the spatial dimension.Results show that this method can well process the temporal features and spatial information.Meanwhile,it receives an average accuracy of 90.63% on the SEED dataset,and average accuracies of 92.09% and 92.82% on the DEAP dataset respectively,which is superior to most of the existing studies.3)A dual-input convolutional network is developed to further improve the classification accuracies in different EEG recognition tasks.Its main ideas can be summarized as follows: First,a three-layer full convolution structure is used to process EEG sequences and capture their spatial-temporal information;meanwhile differential entropy features in five main frequency bands are extracted as another input of the model to receive their frequency information.This model combines the signal-input model of the first work with the feature-input model of the second work,and learns robust task-related representations from the spatial-temporal-spectral dimensions of EEG signals in dual input mode.Here,this paper use the above two public datasets to evaluate the developed model.Results show that it reaches an average accuracy of95.74% on the SEED dataset,and average accuracies of 96.58% and 96.51% on the DEAP dataset respectively,which receives a degree of improvement compared with the existing studies.