Projeto otimizado de redes neurais artificiais para predição da rugosidade em processos de usinagem com a utilização da metodologia de projeto de experimentos
Carregando...
Arquivos
Data
2011-08-09
Autores
Pontes, Fabrício José [UNESP]
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
O presente trabalho oferece contribuições à modelagem da rugosidade da peça em processos de usinagem por meio de redes neurais artificiais. Propõe-se um método para o projeto de redes. Perceptron Multi-Camada (Multi-Layer Percepton, ou MLO) e Função de Base radial Radial Basis Function, ou RBF) otimizadas para a predição da rugosidade da pela (Ra). Desenvolve-se um algoritmo que utiliza de forma hibrida a metodologia do projeto de experimentos por meio das técnicas dos fatoriais completose de Variações Evolucionária em Operações (EVOP). A estratégia adotada é a de utilizar o projeto de experimentos na busca de configurações de rede que favoreçam estatisticamente o desempenho na tarefa de predição. Parâmetro de corte dos processos de usinagem são utilizados como entradas das redes. O erro médio absoluto em porcentagem (MAE %) do decil inferioir das observações de predição para o conjunto de testes é utilizado como medida de desempnho dos modelos. Com o objetivo de validar o métido proposto são empregados casos de treinamento gerados a partir de daods obtidos de trabalhos de literatura e de experimentos de torneamento do aço ABNT 121.13. O método proposto leva á redução significativa do erro de predição da rugosidade nas operações de usinagem estudadas, quando se compara seu desempenho ao apresentado por modelos de regressão, aos resultados relatados pela literatura e ao desempenho de modelos neurais propostos por um pacotecomputacional comercial para otimização de configurações de rede. As redes projetadas segundo o método proposto possuem dispersão dos erros de predição significativamente reduzidos na comparação. Observa-se ainda que rede MLP atingem resultados estatisticamente superior aos obtidos pelas melhores redes RBF
The present work offers some contributions to the area of surface roughness modeling by Artificial Neural Network in machining processes. Ir proposes a method for the project networks of MLP (Multi-Layer Perceptron) and RBF (Radial Basis Function) architectures optimized for prediction of Average Surface Roughness (Ru). The methid is expressed in the format of an algorithm employing two techniques from the DOE (Design of Experiments) methodology: Full factorials and Evolutionary Operations(EVOP). The strategy adopted consists in the sistematic use of DOE in a search for network configurations that benefits performance in roughess prediction. Cutting para meters from machining operations are employed as network inputs. Themean absolute error in percentage (MAE%) of the lower decile of the predictions for the test set is used as a figure of merit for network performance. In order to validate the method, data sets retrieved from literature, as well as results of experiments with AISI/SAE free-machining steel, are employed to form training and test data sets for the networks. The proposed algorithm leads to significant reduction in prediction error for surface roughness when compared to the performance delivred by a regression model, by the networks proposed by the original studies data was borrowed from and when compared models proposed by a software package intend to search automatically for optimal network configurations. In addition, networks designed acording to the proposed algorithm displayed reduced dispersion of prediction error for surface roughness when compared to the performance delivered by a regression model, by the networks proposed by the original studies data was borrowed from and when compared to neural models proposed by a software package intended to searchautomatically for optimal network configurations. In addition, networks designed according to the proposed algorith ... (Complete abstract click electronic access below)
The present work offers some contributions to the area of surface roughness modeling by Artificial Neural Network in machining processes. Ir proposes a method for the project networks of MLP (Multi-Layer Perceptron) and RBF (Radial Basis Function) architectures optimized for prediction of Average Surface Roughness (Ru). The methid is expressed in the format of an algorithm employing two techniques from the DOE (Design of Experiments) methodology: Full factorials and Evolutionary Operations(EVOP). The strategy adopted consists in the sistematic use of DOE in a search for network configurations that benefits performance in roughess prediction. Cutting para meters from machining operations are employed as network inputs. Themean absolute error in percentage (MAE%) of the lower decile of the predictions for the test set is used as a figure of merit for network performance. In order to validate the method, data sets retrieved from literature, as well as results of experiments with AISI/SAE free-machining steel, are employed to form training and test data sets for the networks. The proposed algorithm leads to significant reduction in prediction error for surface roughness when compared to the performance delivred by a regression model, by the networks proposed by the original studies data was borrowed from and when compared models proposed by a software package intend to search automatically for optimal network configurations. In addition, networks designed acording to the proposed algorithm displayed reduced dispersion of prediction error for surface roughness when compared to the performance delivered by a regression model, by the networks proposed by the original studies data was borrowed from and when compared to neural models proposed by a software package intended to searchautomatically for optimal network configurations. In addition, networks designed according to the proposed algorith ... (Complete abstract click electronic access below)
Descrição
Palavras-chave
Redes neurais (Computação), Neural networks
Como citar
PONTES, Fabrício José. Projeto otimizado de redes neurais artificiais para predição da rugosidade em processos de usinagem com a utilização da metodologia de projeto de experimentos. 2011. 259 f. Tese (doutorado) - Universidade Estadual Paulista, Faculdade de Engenharia de Guaratinguetá, 2011.