Dynamic Optimization Design-based Deep Learning Model And Its Application On Eeg-based Analysis

The electroencephalogram(EEG)signal is considered to be one of the most reliable and effective measurements for identifying the physiological state of human and has been widely used by researchers.Deep learning technologies provide new technical approaches for mining information in EEG signals,but also face the problem of structure design and the noise in signals.In this dissertation,we develop swarm intelligence algorithm and reinforcement learning method to realize structure optimization,combining with recurrence networks for input optimization,to realize dynamic optimization design for deep learning models in emotion recognition and fatigue driving detection tasks.In this dissertation,we design the optimization framework based on the EEG data of emotion recognition task.We vectorize the convolutional neural network structure parameters,and use the binary code system to avoid the different mathematical representations of the same model structure.After that,through the development of the particle swarm optimization(PSO)algorithm,with the gradient penalty and other strategies introduced,the gradient particle swarm optimization(GPSO)algorithm is applied to the optimization of the networks model.The classification results illustrate that the model optimized by the GPSO algorithm can achieve valid accuracy on the three-category task.In this dissertation,we analyze the classical artificial features,and combining the coincidence filtering,we propose a function module to realize and enhance the function of these artificial features.Based on this function module,we construct a baseline model CF-CNN to analyze two different EEG signal tasks.The results of two different EEG-based tasks illustrate that the CF-CNN model can achieve valid classification performance on both of the tasks,and has the advantages of high computational efficiency and low training cost.Then,we introduce the reinforcement learning method,combined with the CF-CNN baseline model to achieve structure optimization.The results illustrate that the optimized search method based on reinforcement learning and CF-CNN baseline model can improve the efficiency of structure optimization.In order to reject the influence of low signal-to-noise ratio,in this paper,we use multiplex recurrence networks(MRNs)to process the EEG signals and optimize the input of CNNs.We reconstruct the phase space of each channel's EEG signal,then construct a recurrence network based on the two-norm distance in the phase space,and connect the phase space of different channels through MRNs.By constructing multiplex recurrence networks,we realize the integration of information between different brain regions,and the mutual information matrix is derived as the input of the CNNs.The results illustrate that the MRNs-based CNNs perform better than other traditional methods and advanced methods in fatigue driving detection task.