Artificial neural-network-based control system for springback compensation in press-brake forming

The effect of the training set size and the number of input parameters on the predictive capability of a neural-network-based control system for springback compensation in air bending was investigated. An aluminium alloy, in the form of 2, 3 and 4mm thick sheets, given by two suppliers, was bent to obtain the database for learning. The influence of the number of input parameters on the network performances was studied using input patterns consisting of six and seven parameters; the punch stroke was the output. A neural network for each sheet thickness was built and trained using two training levels; an approach building a unique network, trained using the entire database, was also followed. The results obtained in the operational mode have shown that, if the training patterns are well balanced, the increase in the training set size and in the number of input parameters produces an improvement in the network performances. The networks trained with all data available were characterised by a generalisation capability that is slightly lower than the one for the networks trained using data from sheets with a given thickness.