Improved Extreme Learning Machine For Indentification And Classification Research

With the deep research of neural networks, feed-forward neural networks have some shortcomings, such as slow convergence, difficulty to adjust the parameters and over-fitting puzzle, which affect its further applications. To address the above issues, the thesis focuses on the common static problem in neural network research:identification and classification. Based on the Extreme Learning Machine theory, new methods are proposed for nonlinear system identification and epilepsy EEG signal classification. This thesis makes the following research results:1. The quasi-linear Extreme Learning Machine identification algorithm is proposed.Taking advantage of ARX linear structure and high convergence speed of Extreme Learning Machine, quasi-linear Extreme Learning Machine (QL-ELM) is proposed for nonlinear system identification. Since Huang et al have demonstrated that with increasing number of the hidden layer nodes, a minimum training error can be obtained. However, with the increasing of hidden layer nodes, the output layer weights will decreased gradually. In this case, many nodes having no effect in the final results are called "useless nodes", which increase computational complexity. Therefore, in order to reduce the complexity of neural network, a quasi-linear identification structure is proposed. A macro linear equation expression nested with nonlinear coefficient is used to represent the relationship between inputs and outputs. Linear and nonlinear iterative identification method is used to approximate common nonlinear system. This identification model can compress the number of hidden layer nodes, and simplify the structure of neural network to speed up its convergence speed. Comparative analysis and experiments prove that the proposed identification method has higher identification accuracy.2. An implement of internal model control based on quasi-linear Extreme Learning Machine algorithm.Based on the above proposed quasi-linear method, the quasi-linear model and inverse model controller are established. Due to the high recognition accuracy of proposed QL-ELM method, the control deviation caused by modeling error can be reduced. This method is used in chemical nonlinear CSTR process. Because of the high accuracy of modelling, a better control performance is obtained. Set point control without steady deviation and better robust performance can be achieved.3. A hybrid multiplicative node Extreme Learning Machine classifier is proposed.For the epileptic nonlinear classification problem, a hybrid multiplicative node Extreme Learning Machine (M-ELM) is proposed. Based on the conventioal ELM, multiplication nodes are added in the hidden layer and increase the nonlinear expression ability of the network. A series feature extraction methods, including discrete wavelet transform, phase space reconstruction and singular value decomposition are combined together to extract the feature as inputs of the classifier. The sequential extraction methods are able to extract the EEG signal feature precisely and reduce the information loss and dimension. Hybrid multiplicative ELM classifier increases the hidden layer expression of ELM. The small difference can be mapped to higher dimensional space and achieve higher classification accuracy compared with tradition ELM.Finally, the whole work of this article is summarized and outlook for further study is provided.