Optimized Improvements And Application Of Extreme Learning Machine

With the coming of the Internet Plus Age,the amount of data are growing at an explosive rate,no matter in Internet industry,fast consumer goods industry,financial industry and manufacturing industry.At the same time,there are tasks like distribution,storage,analysis,visualization of big data which are hard challenge for enterprises.As an important method for analyzing and mining hidden patterns and laws,machine learning technology has increasingly exhibited its importance and power.Now,the research and application of machine learning has been widely used in image recognition,speech analysis,natural language processing and data mining.In the machine learning field,classification and regression are two basic problems of the task.Meanwhile,in the research process of machine learning,many general algorithms such as Generalized Linear Model(GLM),Artificial Neural Network(ANN),Support Vector Machine(SVM),and Extreme Learning Machine(ELM)have appeared rapidly.How to improve the prediction accuracy of the classifier or the fitting precision of the regression machine while improving the generalization ability of these models have become an important issue in the development of machine learning technology.Regarding this background,the main contributions of this thesis can be summarized as follows:(1)The weights between the input layer and hidden layer which are randomly generated may cause the low rank problem in the hidden layer output matrix H.it will further lead to the redundancy problem of hidden layer nodes.For solving this problem,a novel method called Correlation Coefficient Mapping-Extreme Learning Machine(CCM-ELM)is proposed in this thesis.It makes use of the correlation coefficients between input features and prediction labels to map the results to a new domain,and determines the final weights between the input layer and hidden layer.This method can improve the prediction accuracy or fitting precision in classification or regression problems,while by making more efficient use of hidden layer nodes and improves the generalization ability of the model.(2)The problem that the homogenization of activation functions in hidden layer standard ELM,a novel method called Mixed Domain-Extreme Learning Machine(MD-ELM)based on particle swam optimization is proposed in this thesis.In this algorithm,the hidden layer activation function(activation function sequence contains 7 kinds of candidates)is defined as a single particle,we randomly generated a large number of particles as the initial population,to find the optimal individual in accordance with certain rules of evolutionary iterations.The proposed algorithm can effectively improve the utilization of the hidden layer nodes,and the generalization ability of the model.(3)In the process of petrochemical production,the flow corrosion in the equipment and pipeline occur frequently.With respect to this,two experiments have been done and their corresponding data are gathered.1)the erosion rate of #10 carbon steel under different experimental conditions was obtained by using the experimental equipment in accordance with the fixed variable method.2)By using CFD simulation software,the erosion rates(including average rate and maximum rate)of 90°elbow under different conditions are simulated.The above two kinds of corrosion were analyzed and modeled using different machine learning methods including those two proposed in this thesis.The comparative results has proved that two proposed methods,improved Extreme Learning Machine(CCM-ELM and MD-ELM)can fit and predict the corrosion data with better performances which could be feasible methods for corrosion prediction problem in petroleum chemical industry.The above research results have certain innovation and practical significance.Based on the theoretical analysis and practical application,this thesis has made some breakthroughs while expand the usage and thoughts of Extreme Learning Machine in machine learning field and industrial applications.