A study on the performance of various predictive models based on artificial neural network for backward metal flow forming process
Authors :- Prabas Banerjee, Rupam Laha, Mithilesh K. Dikshit, et al.
Publication :- International Journal on Interactive Design and Manufacturing (IJIDeM) (2022)
Backward metal flow forming is an incremental forming process that is used to manufacture near shape products with precise dimensions. Among the most important application fields are defence, aerospace, and automobiles. The incremental forming method is used in this study to produce large tubes with precise wall thickness. These are the inner shells that contain the rocket fuel. Even though the experimental setup is simple, the localized type of deformation in this forming process makes it difficult. Estimating the final dimensions of the produced tubes can be challenge due to a lack of literature on the relationship between the input and output parameters. In this research, two artificial neural network architectures are used that are trained with three optimization techniques (gradient descent, Broyden–Fletcher–Goldfarb–Shanno, and Levenberg–Marquardt) respectively. A comparative analysis based on statistical techniques are done to identify whether the predictive models are different from each other. As, the bias values on responses are correlated and a multivariate analysis of these correlated bias values were compared using a multivariate analysis of variance. The statistical analysis showed that all the models were equally effective in the prediction of springback and ovality, although significant variance was observed for inside diameter. It is concluded that the BFGS tuned Elman neural network delivers satisfactory statistical performance for all three output factors simultaneously.