EFFECT OF THE TRAINING SET SIZE ON SPRINGBACK CONTROL BY NEURAL-NETWORK IN AN AIR BENDING PROCESS

The present investigation focused on the development of an intelligent air bending process using a neural network based control system. The emphasis was on the effect of the training set size on the predictive performances of the neural networks. The AA 5754 aluminium alloy, in f orm of 2, 3 and 4 mm thick sheets, provided by different producers, wa s bent to obtain the data base for training. Each input pattern was co nstituted by the bend angle after unloading, sheet thickness, and five parameters describing the mechanical behaviour of the material derive d by a model using 'in-process' measurements of bending force and punc h displacement; the punch stroke was the unique output pattern. Three neural networks, one for each sheet thickness, were built and trained using 30 and 60 input/output patterns, respectively. In the operationa l mode, the sheet thickness and desired bend angle were the external n etwork inputs whilst the input parameters describing the mechanical be haviour were estimated 'in-process'; the punch displacement required t o compensate springback was then calculated by the neural networks. Th e 2 and 3 mm thick sheet networks show an increase in the predictive p erformances with increasing the training set size whilst a very slight decrease is observed for the 4 mm thick sheet network. (C) 1998 Elsev ier Science S.A. All rights reserved.