基于神经元特性的径向基函数神经网络自组织设计方法

Design of Radial basis function neural network based on neuron characteristicsDesign of Radial basis function neural network based on neuron characteristics

中文关键词  径向基函数神经网络  自组织  结构设计  二阶算法  非线性系统建模

英文关键词  Radial basis function neural network  self-organization  structural design  activity  significance  correlation  second-order algorithm

中文摘要

      针对径向基神经函数(RBF)网络隐层结构难以确定的问题, 本文介绍了一种基于神经元特性的RBF神经网 络自组织设计方法, 该方法将神经元的激活活性、显著性、相关性相结合设计RBF(ASC–RBF)神经网络. 首先利用 神经元的激活活性, 实现隐含层神经元的自适应增加, 结合神经元的显著性以及神经元之间的相关性, 实现神经元 的自适应替换和合并, 完成网络自组织设计并提高网络的紧凑性, 然后利用二阶梯度算法对网络参数进行修正学 习, 保证了RBF网络的精度; 另外, 针对网络结构自组织机制给出了稳定性分析; 最后通过两个基准非线性系统建模 仿真实验以及实际污水处理过程水质参数预测实验验证, 证明该算法的有效性. 对比实验结果表明, ASC–RBF神经 网络与现有的自组织网络相比, 在保证泛化性能的同时, 该网络的训练速度更快, 而且有更紧凑的网络结构.

英文摘要

      Aiming at the problem that the hidden layer structure of radial basis neural function (RBF) neural network is difficult to determine, this paper introduces a self-organizing design method of RBF neural network based on the characteristics of neurons. This method combines the activation activity, significance and correlation of neurons Combined design of RBF (ASC-RBF) neural network. Firstly, The network uses the activity of neurons to adaptively increase the hidden layer neurons, and combines with its significance and correlation to complete the adaptive replacement and merging of neurons. Furthermore, the self-organizing design of the neural network is completed and its compactness is improved. Then, a second-order algorithm is used to modify the network parameters to ensure the accuracy of the RBF network. In addition, a stability analysis is given for the network structure self-organization mechanism. Finally, in order to verify the effectiveness of the proposed ASC-RBF network, two benchmark nonlinear system modeling experiments and a water quality parameter prediction experiment in a wastewater treatment system are performed. The results demonstrate that compared with the existing self-organizing network, the ASC-RBF neural network has faster training speed and a more compact network structure while ensuring generalization performance.