Comparison between regression and artificial neural network for prediction model of flexibly reconfigurable roll forming process

Abstract

Flexibly reconfigurable roll forming (FRRF) is a novel sheet metal forming technology in which sheet metal is shaped into the desired curvature using reconfigurable rollers and gaps. However, FRRF has a disadvantage in terms of predicting the forming result because this technology produces two-dimensional curved surfaces from one-dimensional curved lines. A previous study used regression analysis to predict the longitudinal curvature results of the FRRF process. In this study, a prediction model using machine learning was developed. A comparative investigation with an existing regression model was conducted using an artificial neural network (ANN) model based on machine learning. Sample data were obtained via numerical simulation, and the target surfaces were convex and saddle-shaped. To confirm the validity of each model, a goodness-of-fit test was performed. Further, to confirm the predictability of each model, the same amount of random data was extracted and a random sample test was performed. In the case of the saddle-shaped surface, the ANN model showed a predominantly high predictability, but in the case of the convex shape surface, the regression model performed better. The ANN model has potential to be used in practical applications such as skin structure forming for aircraft and shipbuilding.