Redes neurais artificiais e planejamento fatorial:um estudo da integridade superficial do aço ABNT 52100 endurecido no processo de torneamento
Carregando...
Arquivos
Citações na Scopus
Tipo de produção
Dissertação
Data
2012
Autores
Paschoalinoto, N. W.
Orientador
Delijaicov, Sergio
Periódico
Título da Revista
ISSN da Revista
Título de Volume
Citação
Texto completo (DOI)
Palavras-chave
Redes neurais (Computação),Tensões residuais
Resumo
O uso de redes neurais artificiais e planejamento fatorial, para o estudo da integridade superficial de peças de aço ABNT 52100 torneadas no estado endurecido, foram o foco deste trabalho. Os ensaios de torneamento foram efetuados de acordo com um planejamento experimental, utilizando-se três insertos distintos: Tipo-S (20o e 30o) e Tipo-T. Os testes foram executados em 60 amostras de aço ABNT 52100 com dureza média de 58,5 HRC. Os dados de corte (velocidade de corte, avanço e profundidade de penetração) utilizados no torneamento dos corpos de prova foram definidos como variáveis independentes da rede neural artificial e do planejamento fatorial e as forças de usinagem (força de avanço, força de corte e força de penetração) foram medidas com um transdutor tri-axial e definidas como variáveis dependentes. Após a usinagem mediram-se as rugosidades Ra e as tensões residuais circunferenciais, também nomeadas variáveis dependentes. As tensões residuais foram medidas pelo método da difração de raios-X. A análise metalográfica em todos os corpos de prova permitiu verificar a presença ou não da camada branca decorrente do processo da usinagem. Algumas configurações de redes neurais artificiais foram sugeridas, treinadas e testadas. Obteve-se grande aproximação entre os resultados experimentais e os previstos pelas redes. Um planejamento fatorial completo de três variáveis e dois níveis com dois pontos centrais e uma réplica foi utilizado para a análise estatística. Os resultados preditos pelas redes neurais propostas foram comparados com os resultados estatísticos fatoriais através da análise das respectivas variâncias. A otimização das respostas, predizendo a integridade superficial, em função das variáveis independentes para cada inserto foram obtidos por meio da função desirability e uso do programa Statistica. Os resultados mais significativos foram: as redes neurais propostas e o planejamento estatístico predizem as respostas de forma semelhante; o avanço foi a variável mais importante para a rugosidade; a velocidade de corte e a profundidade de penetração influenciaram significativamente as tensões residuais para o inserto do tipo S e o avanço e a profundidade de penetração para o inserto T. Constatou-se que o inserto tipo S com 30º favoreceu a formação da camada branca.
The use of artificial neural networks and factorial plan for the surface integrity study of ABNT 52100 steel pieces which are turned in the hardened condition, were the focus of this work. Turning tests were carried out according to an experimental planning, using three different inserts: S-Type (20º and 30º) and T-Type. The tests were performed in 60 samples of ABNT 52100 steel with average hardness of 58,5 HRC. The cutting conditions (feed rate, cutting speed and depth of cut) used in the turning of the samples were defined as independent variables of the artificial neural network and the factorial plan and the machining forces (feed force, cutting force and passive force) were measured with a tri-axial transducer and defined as dependent variables. After the machining operation, the Ra roughness and the circumferential residual stress, also named dependent variables, was measured. The residual stresses were measured by the X-Rays diffraction method. The metallographic analysis of all samples allowed to verify the presence or not of the white layer resulting from the machining process. Some configurations of artificial neural networks were suggested, trained and tested. A good approximation between the desired and the planned results by the networks was obtained. A complete factorial plan of three variables and two levels with two central points and a replica were used for this statistical analysis. The results predicted by the proposed neural networks were compared with the factorial statistical results through the respective variances analysis. The responses, optimization predicting the surface integrity, in function of the independent variables for each insert were obtained by the desirability function and the use of the Statistica software. The most significant results were: the proposed neural networks and the statistical plan predicted the answers in a similar way; the feed rate was the most important variable for roughness; the cutting force and the passive force influenced significantly the residual stress with the insert of the S type and the feed and the passive force with the T insert. It was found that the S type insert with 30 ° favored the white layer formation.
The use of artificial neural networks and factorial plan for the surface integrity study of ABNT 52100 steel pieces which are turned in the hardened condition, were the focus of this work. Turning tests were carried out according to an experimental planning, using three different inserts: S-Type (20º and 30º) and T-Type. The tests were performed in 60 samples of ABNT 52100 steel with average hardness of 58,5 HRC. The cutting conditions (feed rate, cutting speed and depth of cut) used in the turning of the samples were defined as independent variables of the artificial neural network and the factorial plan and the machining forces (feed force, cutting force and passive force) were measured with a tri-axial transducer and defined as dependent variables. After the machining operation, the Ra roughness and the circumferential residual stress, also named dependent variables, was measured. The residual stresses were measured by the X-Rays diffraction method. The metallographic analysis of all samples allowed to verify the presence or not of the white layer resulting from the machining process. Some configurations of artificial neural networks were suggested, trained and tested. A good approximation between the desired and the planned results by the networks was obtained. A complete factorial plan of three variables and two levels with two central points and a replica were used for this statistical analysis. The results predicted by the proposed neural networks were compared with the factorial statistical results through the respective variances analysis. The responses, optimization predicting the surface integrity, in function of the independent variables for each insert were obtained by the desirability function and the use of the Statistica software. The most significant results were: the proposed neural networks and the statistical plan predicted the answers in a similar way; the feed rate was the most important variable for roughness; the cutting force and the passive force influenced significantly the residual stress with the insert of the S type and the feed and the passive force with the T insert. It was found that the S type insert with 30 ° favored the white layer formation.