One of the greatest challenges of manufacturing sheet metal parts is to obtain consistent part dimensions.  Springback, the elastic strain recovery in the material when the tooling is removed, is the major root cause of variations and inconsistencies in the final part geometry.  Obtaining a consistent and desirable amount of springback is extremely difficult due to the non-linear effects and interactions from process and material parameters.  In this work, the exceptional ability of a neural network along with a stepped binder force trajectory to control springback angle and maximum strain is demonstrated for a simulated channel forming process.  When faced with even large variations in material properties, sheet thickness, and friction coefficient, our control system produces a robust final part shape.

        Neural networks must be first trained to learn the relationship between the inputs and outputs.  Here the inputs were four polynomial curve fit coefficients from the punch force trajectory, and the outputs were the high binder force (HBF) and percentage of the punch displacement (PD%) where the stepped binder force should occur.  Trial and error numerical simulations were conducted with variations in the material properties, sheet thickness, and friction coefficient to create training examples from which the network could extract the input-output pattern.  Then, the backpropagation algorithm was used to train the network, and other cases of varying material properties, sheet thickness, and friction coefficient were feedforward in the trained network to obtain the stepped binder force parameters.  Finally, a comparison with a closed-loop control system was investigated.
        The results of this neural network control system were springback angles between 0.2 to 0.6 degrees and maximum principal strain values of 8% to 10% for even large variations in material properties, sheet thickness, and friction coefficient.  When compared with a closed-loop control system of the optimal punch force trajectory, the neural network produced more accurate values of springback angle and maximum strain as well as being a more robust process and simpler to implement. 

        As a follow up to the simulated aluminum channel forming research, experimental implementation of the neural network with a stepped binder force trajectory for a steel channel forming process has been completed.  Again, the neural network was able to predict the HBF and PD% in order to control the springback within a reasonable range even for cases which the neural network was not trained with.

• Vikram Viswanathan, Brad Kinsey, and Jian Cao (2003) “Experimental Implementation of Neural Network Springback Control for Sheet Metal Forming”, Journal of Engineering Materials and Technology, Vol. 125, No. 2, pp. 141-7.
• Jian Cao, Brad Kinsey, and Sara Solla (2000) “Consistent and Minimal Springback Using a Stepped Binder Force Trajectory and Neural Network Control”, Journal of Engineering Materials and Technology, Vol. 122, No. 1, pp. 113-8.