Análise de métodos de agrupamento para o treinamento de redes neurais de base radical aplicadas à identificação de sistemas

AUTOR(ES)
DATA DE PUBLICAÇÃO

2006

RESUMO

For complex systems, modeling using basic laws to determine their dynamic behavior is not always possible. An alternative to solve these problems is use concepts of systems identification. Trough system identification it is possible to determine a mathematical model based on input and output data of the system. When little prior knowledge is available, it is common to use a black-box mathematical model to represent different nonlinear systems. The neural network models have proven to be successful nonlinear black-box model structures in many applications. A conception of neural network that can be applied to the systems identification is the radial basis function. This network follows a local approximation approach and is composed by a hidden layer that is defined by a set of radial basis functions. The hidden units supply a set of radial basis functions that constitute an arbitrary base for the patterns of input. In order to train a radial basis neural network, the clustering algorithms are applied for determination of the centers of each radial basis function, aiming to discover patterns in the input data. The objective of this dissertation is the analysis and comparison of k-means, fuzzy c-means, Gustafson-Kessel and Gath-Gheva clustering methods in the radial basis functions neural networks, applied in nonlinear identification of a heat exchanger, Mackey-Glass system, Rössler chaotic system and Box-Jenkins gas furnace.

ASSUNTO(S)

engenharia de producao redes neurais (computação) algoritmos sistemas não-lineares - identificação

ACESSO AO ARTIGO

Documentos Relacionados