Research And Application Of Two Interval Neural Networks

Interval neural networks(INNs)aims to provide an effective method for treat uncertain data in the field of imprecise data modeling.This learner model uses interval to represent the uncertain data and take advantage of neural network to finish modeling.Interval feed-forward neural networks with error back-propagation algorithms(IBPNNs)as a popular kind of INNs,catches many attention in recent years.However,the gradient-based learning algorithms for training IBPNNs suffer from local minima problem,slow convergence and very poor sensitivity to learning rate setting.And the intelligent optimization learning algorithm is a time consuming process experience to convergence.The problem of both learning algorithm limit the application of IBPNN.To overcome these difficulties,this thesis proposed two new INNs—interval RBF neural network(IRBFNN)and interval RVFL network(IRVFLN).Moreover,we applied the INNs into the complex industrial process,which laid the foundation for region optimization(interval optimization).The thesis includes the following aspects:Firstly,we construst the IRBFNN model.Both the structure and learning algorithm are given in this thesis.And the learner model is complete INN,which using subtractive clustering algorithm to calculate the parameters of hidden layer and adopt BP algorithm to tune weights of output layer.The simulation results show that the convenging and the approximating ability of IRBFNN are both better than IBPNN.Secondly,we extend the random vector function-link(RVFL)networks with single-hidden-layer to interval ones(IRVFLNs)with interval model parameters.And according to the from of network parameters,IRVFLNs are dividen into two types.The same characters of both interval learner model are that the weights of hidden layer can be randomly assigned and no need to be tuned and output weights are determined by solving linear equation systems.Results indicate that IRVFLN-I outperforms IRVFLN-II and IBPNN.Thirdly,as an application case study,the IRVFLN is employed to model the glutamic acid fermentation process.Taking consideration of the measuring errors of sensors,the manipulated variables and the state variables of the fermentation process are expanded to interval values,which been employed to train IRVFLN to set up the process model.Through introducing the similarity quantitative function,to culster the fermentation process that have different initial condition,and construct the enaemble interval RVFL network(EIRVFLN)for glutamic acid fermentation process The testing results indicate that the EIRVFLN model with the interval outputs not only can produce the prediction values of the state variables,but also show the forecasting errors.The accuracy of the EIRVFLN model meets the manufacture requirements through the application results analysis.At last,the thesis also presented the next improving direction based on the above research results.