Research On Radial Basis Neural Network Structure Design And Online Modeling Method

Radial basis function neural networks have been widely used in system modeling owing to their simple structure,fast convergence,and strong nonlinear mapping ability.Actual systems usually have time-varying and uncertain characteristics,which require the established system models to be adaptable to new environments,making it an inevitable trend to study online modeling methods for RBF neural networks.However,how to determine the structure of the network and to adjust the structure and parameters of the network in real time according to the online data and improve the online self-adaptive capability of the network are the present difficulties in the research of online modeling methods.Consequently,this thesis focuses on RBF neural networks,and investigates the structure design and online modeling methods of two RBF neural network models,feedforward and recurrent,respectively,and applies them to the online detection of key water quality parameters in the wastewater treatment process.The main research work and innovations of the thesis are as follows:1.Research on structure design and online modeling method of feedforward RBF neural networkFocusing on feedforward RBF neural network,the adaptive initialization algorithm of RBF neural network is studied to automatically determine the structure and parameters of the network,and then combined with the sliding window algorithm,the online self-adjustment algorithm of parameters is designed to realize the online modeling of the network.The self-organization algorithm of RBF neural network structure is investigated to make the structure and parameters of the network realize adaptive adjustment according to the online condition,which further improves the online modeling ability of the network.(1)Online self-adjusting RBF neural network with parameters based on micro-integer tuning algorithm1)RBF neural network adaptive initialization algorithmFor initializing the RBF neural network and determining the structure and initial parameters of the network,an improved Canopy-K-means algorithm is proposed in this thesis.Firstly,a rough clustering of the data is performed by using the density-based Canopy algorithm.Secondly,the appropriate number of clusters and initial central parameters are determined adaptively utilizing the maximum weight product method.Finally,the results are used as initial parameters of the K-means clustering algorithm for learning the RBF network.The experimental results show that the proposed algorithm can automatically determine the optimal structure of the network,and the network can obtain better classification accuracy and function approximation accuracy.2)Online self-adjusting algorithm for RBF neural network parametersBased on the RBF network initialization algorithm,an online self-adjusting RBF neural network(OA-RBFNN)with parameters based on the micro-integer tuning algorithm is designed in combination with the sliding window mechanism.Firstly,the network structure and its initial parameters are determined using the network initialization method.Secondly,based on the changes of samples in sliding windows,the clustering-based fine-tuning algorithm combined with the gradient-based integer-tuning algorithm is used to learn the parameters of the network,realizing online self-adjustment of network parameters.The experimental results have shown that the proposed OA-RBFNN has good generalization performance.(2)Online self-organizing RBF neural network based on Gaussian MembershipTo further improve the online adaptive capability of the network and reduce the sensitivity of the network to hyperparameters,an online self-organizing RBF neural network based on Gaussian Membership(GM-OSRBFNN)is investigated in this thesis.Firstly,Gaussian Membership is introduced to enhance the insensitivity of the network to hyperparameters,as well as to represent the similarity between hidden layer neurons and samples and between neurons and neurons as a similarity metric.Secondly,Secondly,the similarity metric is used to design neuron addition and merging rules for self-organization of the network structure,while error constraints are introduced in the neuron adding rules and noise neuron deleting rules are designed to make the network structure more compact.Finally,the online fixed small batch gradient algorithm is proposed for online learning of parameters,which ensures the fast and stable convergence of the network.The experimental results demonstrated that the proposed GM-OSRBFNN model can achieve online self-organization of network structure,and the network has a compact structure and high generalization performance.2.Research on structure design and online modeling algorithm of recurrent RBF neural networkFocusing on recurrent RBF neural networks,this thesis investigates the recurrent mechanism of recurrent RBF neural networks and designs an input-output directly connected recurrent RBF neural network with memory factor.An adaptive batch online learning algorithm is proposed to realize online modeling of recurrent RBF networks.(1)An input-output directly connected recurrent RBF neural network with memory factorIn view of the problem that the recurrent mechanism of traditional recurrent RBF neural network causes the decay of historical memory,which affects the network performance,an input-output directly connected recurrent RBF neural network structure with memory factor(MF-DIOCRRBF)is designed.Firstly,a memory factor is introduced into the recurrent mechanism of the hidden layer neurons to accumulate the historical information selectively and provide a more historical information to the network.Secondly,the input-output direct connection structure is added to the input and output layers of the network to improve the performance of the network.Finally,a gradient algorithm with adaptive learning rate is designed to learn the network using Lyapunov method,which ensures fast and stable convergence of the network.The experimental results show that the proposed MF-DIOCRRBF structure has a fast convergence speed and good prediction accuracy.(2)An adaptive batch online learning algorithm for recurrent RBF neural networksAiming to improve the online learning capability of recurrent RBF networks,a recurrent RBF neural network adaptive batch online learning algorithm is proposed.Firstly,the online sample acquiring module is designed based on the sliding window strategy.Secondly,the gradient algorithm with adaptive learning rate is developed to adjust the model parameters online,which ensures stable and fast convergence of the model within a single sliding window.Finally,the batch adaptive adjustment algorithm is proposed to adjust the batch size according to the network performance,improving the prediction accuracy of the network.The experimental results demonstrated that the proposed adaptive batch online learning algorithm enabled the online modeling of recurrent RBF networks,while improving the online modeling accuracy of the networks.3.Online monitoring model of effluent ammonia nitrogen based on RBF neural networkTo verify the effectiveness of the RBF neural network structure design and online modeling method investigated in this thesis,the proposed feedforward and recurrent RBF network online models are applied to the online monitoring problem of key water quality in the wastewater treatment process.Firstly,for the problem of difficult real-time and online measurement of effluent ammonia nitrogen concentration,an online soft measurement model of RBF neural network based on online structural self-organization algorithm is proposed,which realizes fast and accurate online measurement of ammonia nitrogen concentration in the wastewater treatment processes.Secondly,to enable accurate online prediction of ammonia nitrogen concentration in the effluent at future moments during the wastewater treatment processes,a recurrent RBF neural network online prediction model based on adaptive batch algorithm is proposed.