Research On Online Self-organization Design Method Of Modular Neural Network And Its Applications

Artificial neural network（ANN）is an intelligent computing system,which imitates the information processing of human brain.ANN has widely used for nonlinear system modeling due to its self-learning and good nonlinear approximation ability.However,in practice,many actual processes are the complex nonlinear dynamic processes with the characteristics of complexity,nonlinear,uncertainty,and time-varying.Single neural network has the catastrophic forgetting problem and can’t learn all the samples effectively in learning process.Therefore,the nonlinear dynamic system modeling and prediction is one hot but difficult issue in the study of neural network.Modular neural network is inspired by the modular characteristics of brain,which can effectively improve the computational performance and alleviate the“catastrophic forgetting” of neural network.Therefore,modular neural network is currently the research focus for nonlinear dynamic system modeling.At present,how to divide tasks effectively and online according to the correlation of the samples or task characteristics and how to dynamically adjust the structure of the subnetwork based on the subtasks to obtain a suitable network structure are the difficult issues for modular neural network.Therefore,For the nonlinear dynamic system modeling and time series prediction,this thesis has carried out systematic and in-depth research on the design of online self-organizing method for modular neural network,and the main research work and innovation points are as follows:1.Research on online self-organizing method of modular neural network for nonlinear systems1).An online self-organizing feedforward neural network is proposed to solve the problem that subnetworks of modular neural network are usually the feedforward neural network with the fixed structure,and the initial structure of subnetwork is hard to be determined.Firstly,the proposed model grows its structure based on the network error to ensure the learning ability of network.Secondly,local sensitivity analysis method is used to calculate the contribution of hidden layer neuronds and the neurons with small contribution are pruned to ensure the compactness of network structure.Finally,in order to improve the convergence speed and accuracy of the network,the sliding window mechanism is used to update the network parameters.Compared with models,the results show that the developed model can dynamically adjust its structure and has better generalization performance.2).To solve the problem that task decomposition algorithms based on clustering method usually adopts recursive iteration to only update clustering centers,which take no account of the correlation between samples,an online task decomposition algorithm based on distance and density is proposed.This alrorithm is based on the principle that clustering centers have high local densityand the distance between clustering centers is large,which can divide the sample space dynamically by updating the local density of samples continuously to ensure that modular neural network can effectively obtain the local informance of sample to adjust its structure online.3).For the nonlinear dynamic system modeling,an online self-organizing modular neural network model is proposed.Firstly,the clustering algorithm based on distance and density is used to online divide the sample space and dynamically add or prune subnetwork modules.Secondly,network error and local sensitivity analysis method are used to dynamically adjust subnetworks’ structure.Finally,the results of subnetworks are integrated.The simulation results show that the proposed model does not only adjust subnetworks’ structure online,but subnetworks can adaptively adjust its structure and has compact structure and good generalization performance.2.Design for modular neural network based on EMD technology for time series prediction1).To solve the problem that modular neural network divides time series based on the sample space,which takes no account of the characteristics of time series,an adaptive modular neural network model based on empirical mode decomposition（EMD）is proposed.Firstly,EMD method is used to decompose time series into serveal simple and independent subseries.Secondly,sample entropy and euclidean distance are used to calculate the complexity and similarity of subseries and merge the subseries with simpler and large similarity.Finally,the results of subnetworks are integrated with linear weighted sum.The simulation results show that the proposed model can effectively improve the prediction performance of modular neural network.2).To solve the problem that the models based on EMD need the global samples,which cannot deal with the online problems,an online modular neural network based on the sliding window mechanism and EMD is proposed.This model can online decompose time series in silding window,and uses sample entropy and euclidean distance to dynamically assign subseries to subnetworks.Meanwhile,the overlap of subseries is pruned to alleviate the end effect of subseries at different times.The simulation results show that the proposed model can effectively online decompose time series and has good prediction accuracy.3).A modular neural network model based on EMD and multi-view learning strategy is proposed to slove the problem that the models based on EMD will generate sub-series with high frequency when decomposing time series,which may increases the difficulty of prediction for subnetwork.Firstly,EMD method is used to decompose time series with silding window into serveal simple and independent subseries.Secondly,sample entropy and euclidean distance are used to assign subseries to subnetworks.Finally,the sub-series with high frequency are assigned to the modules with multi-view learning strategy and the sub-series with low frequency are assigned to subnetworks with simple structure.The simulation results show that the model based on EMD and multi-view learning strategy can further improve the prediction performance of modular neural network.3.Online prediction model of key effluent water quality parameters in the process of urban sewage treatmentA prediction model based on MNN with EMD and multi-view learning strategy is proposed to deal with the problem that some key effluent parameters in wastewater treatment process are difficult to measure and predict accurately and in real time.Firstly,the silding window mechanism and EMD method are used to online decompose effluent ammonia nitrogen series.Secondly,sample entropy and euclidean distance are used to assign the sub-series with high frequency the modules with multi-view learning strategy and the sub-series with low frequency to subnetworks with simple structure.Finally,the results of the modules with multi-view learning strategy and subnetworks are integrated.The experimental results show that the proposed prediction model can predict effluent ammonia nitrogen in the wastewater treatment process accurately and in real time.