Neural Network Modeling Of Nonlinear System Based On Extreme Learning Machine

Complex nonlinear system is a common type in industrial system.Through the method of neural network modeling of industrial system,we can better understand the system,improve the operation efficiency of industrial system,and reduce the loss rate of industrial process.These have important economic and social value.There are some problems in the modeling of complex nonlinear industrial systems by using neural networks.For example,the learning ability of neural network is insufficient,it is difficult to identify large unknown systems,and the identification accuracy is low.In view of the above problem,it is difficult to establish an efficient modeling method with universal applicability for a class of nonlinear dynamic systems in complex industries.A new method,intelligent modeling method of alternating group identification,is proposed.The internal network structure of the new structure is optimized.The whole identification system is based on feedforward neural network.Therefore,this thesis first addresses the problem of slow learning in feedforward network when dealing with large data sets and large network embedded systems.By analyzing and discussing the selection of norm and smooth approximation function,a new regularization method of group L1/2 is established.This method uses a new approximation function based on lower bound smoothness.The lower bound smooth approximation function can effectively solve the problems existing in the regularization process,such as the non-smooth problem and the resulting oscillation phenomenon.The redundant dimensions in feedforward neural networks and their corresponding input nodes are filtered.Thus,the sparse processing of input layer of feedforward network is achieved.The learning efficiency and generalization ability of neural network will be improved.The building part of the model includes two parts,the choice of single model and the establishment of alternate identification module.The choice of single model is an important part of combination identification.The selection method of traditional single model is generally based on the weight of linear combination of empirical selection,and the efficiency and accuracy of this choice need to be improved.To solve this problem,this thesis proposes a selection strategy based on the above regularization method,namely the lower bound smooth group L_1/2regularization.The new selection strategy can avoid collinear problem of selected single item model.Furthermore,the selection strategy can be used for model selection in a more reasonable and efficient way.Secondly,an alternate identification module is established.In order to solve the problem of nonlinear modeling in complex unknown system modeling,an alternate identification module based on forgetting factor least square method and extreme learning machine is proposed.Firstly,a general system of nonlinear discrete systems with universality is established.Then the high and low order separation and rolling optimization identification of the system are carried out.That is,the high order identification part of the extreme learning machine is calculated by the low order recursive least squares identification.Finally,feedback and secondary identification are carried out to achieve the purpose of complex system intelligent modeling.This method can overcome modeling errors and structural uncertainties of controlled objects.Thus,the identification of complex systems is more organized and simplified,and the identification process is faster and more accurate.At the same time,extreme learning machine hidden layer node randomness is easy to produce a large number of redundant nodes.In order to solve this problem,the extreme learning machine is regularized by the selection strategy of lower bound smooth group L_1/2.It can make the extreme learning machine reach the"sparsest"state,so as to improve the identification efficiency of the system and prevent the overfitting of learning results.Finally,the effectiveness and universality of the intelligent modeling method for alternate group identification are proved by multi-party experiments.