An Improved Wavelet Neural Network Structure Optimization Algorithm And Application Research In EEG Classification

The Wavelet Neural Network(WNN) is a kind of Artificial Neural Network based on Wavelet Theory. Wavelet Neural Network has been widely used in the fields of nonlinear function approximation, signal classification, dynamic modeling, diagnosis and prediction, and so on. For example, Zhanyong Wang combines Wavelet Neural Network and Genetic Algorithm for fine scale estimation of carbon monoxide concentration and fine particulate concentration. The researchers have shown that the design of WNN network structure plays a vital role in the performance of network. In the training process of WNN, the redundant or simple hidden layer nodes affect the training effect of WNN. In addition, as the WNN like BP uses the gradient descent method to train WNN network in the learning process, learning process is prone to the problem of slow convergence rate, the existence of shock and easy to fall into the local optimal situation. Otherwise, the choice of WNN motivation function also directly affects the accuracy of network training.In order to solve the above questions, we proposed an improved wavelet Neural Network structure optimization algorithm. By using the structure optimization algorithm that we propose in this paper, the network performance of WNN is been improved. In addition, we also create a new EEG data classification model, and apply itto the classification of the EEG data. This model is based on improved wavelet neural network structure optimization algorithm, and it is combined with the brain electrophysiological data(EEG) feature extraction method. Our work of this paper includes:(1) An improved Wavelet Neural Network structure optimization algorithm is been proposed. Firstly, it is to select an incentive function. On the basic of Wang Gaige and other researchers, we construct the wavelet base function warehouse, and then select the most suitable function as the base function of the Wavelet Neural Network. Secondly, the determination of the hidden layer nodes. A new method for determining the initial hidden layer nodes is been proposed. Then we use it determine the hidden layer nodes and update the method of wavelet neural network, to find the appropriate number of hidden layer nodes as soon as possible. Finally, the method of improving the additional momentum and variable learning rate are used to speed up the convergence process in the process of training. The experiment results show that the improved WNN structure optimization algorithm can improve the network performance of WNN. The algorithm can select an appropriate excitation function quickly, and find the effective appropriate hidden layer nodes quickly. Furthermore, the convergence speed can be further accelerated, and the smaller error can be obtained in the test set.(2) A new classification model of EEG data in EEG is also been proposed. Based on the characteristics of EEG data, an improved WNN algorithm is been proposed in this paper, and the GA_HM feature extraction method is been proposed. Firstly, the GA_HM feature extraction method is used to extract the feature of EEG data. Then, we use the improved wavelet neural network structure optimization algorithm to classify the EEG data. In order to verify the effect of GA HM method and improved WNN algorithm, we use the proposed model in this paper to classify the EEG data of 23 subjects. Then, compared with traditional classification algorithms and feature extraction methods, results show that the new EEG data analysis model has higher accuracy rate.