Principal Components Estimation Of Extreme Learning Machine And Its Application Research

Extreme learning machine is a new kind of single hidden layer feedforward neural networks.It has attracted many scholars to do research and application because of its simple structure,good generalization performance and fast learning speed.But when the columns of hidden layer design matrix exists multicollinearity,the hidden layer design matrix becomes non-full rank matrix.Then the output weight calculated by least squares estimation is lack of robustness.To solve this problem,a new ELM algorithm base on principal components estimation called principal components ELM is proposed in this paper.The calculation of output weight by least squares estimation is replaced by principal components estimation.This new ELM algorithm can improve the calculation robustness of ELM.This paper focuses on the following several aspects of research:(1)Firstly,this paper introduces the background and development of extreme learning machine.Then this paper summarizes the research status of extreme learning machine at home and abroad in recent years.The theory of extreme learning machine is introduced and analyzed,mainly analysis the defects and deficiencies of extreme learning machine in this paper.(2)Secondly,this paper expounds the realization process of least squares estimation and problems existed in least squares estimation.Then this paper introduced M estimation,ridge estimation and principal components estimation in robust estimation theory and analyzes their respective characteristics.(3)Thirdly,when the design matrix of extreme learning machine exists multicollinearity,the output weights calculated by least squares estimation would be unstable.In order to solve this problem,this paper introduced the principal components estimation.Combining with the two theories,we proposed a novel approach called PC-ELM,which adopted the principal components estimation method to calculate the network output weights.Thus can effective reduce the effect by multicollinearity existing in the hidden layer design matrix.(4)Fourthly,in order to verify the validity of PC-ELM,a lot of regression and classification experiments compared among several popular ELM learning algorithm.The experimental results verified that PC-ELM keeps the advantages such as fast learning speed,simple structure and good generalization performance.At the same time,PC-ELM can effectively improve the robustness of calculation of output weight in ELM when design matrix is ill.(5)Fifthly,we introduced the process of extracting alumina from coal ash and the process characteristics of alumina rotary kiln.A new method based on PC-ELM to model the firing temperature is proposed in this paper.The training samples are collected from expert control system.PC-ELM is conducted as training algorithm for network model.The experimental results compared with ELM and BP algorithm showed that PC-ELM could get better performance and the forecasting temperature is more accurate than ELM and BP.