Using Neural Network for Springback Minimization in a Channel Forming Process

Springback, the geometric difference between the loaded and unloaded configurations, is affected by many factors, such as material properties, sheet thickness, lubrication conditions, tooling geometry and processing parameters. It is extremely difficult to develop an analytical model for springback control including all these factors. The proposed neural network model is an attempt to deal with such a complicated non-linear system in a predictive way. For demonstration, an aluminum channel forming process is considered in this work. Our previous research [1] has shown that a variable binder force history during the forming operation can reduce the springback amount significantly while maintaining a relatively low maximum strain if an initial low binder force was used followed by a higher binder force. However, when and how much of the increase is depends on the forming conditions of the current process. Here, several numerical simulations using Finite Element Method (FEM) were performed to obtain the teaching data required for training the neural network by means of the back-propagation algorithm. In the predictive mode, different process inputs from the ones used in the previous stage were considered. For each case, the displacement where binder force increases and the level of the high binder force were predicted by the learned neural network and were numerically tested. Consistent low springback angle (< 0.5°) and moderate stretching amount (< 16%) were obtained even in the cases where the process parameters were varied as much as ±25% on friction coefficient and sheet thickness or ±10% on material’s mechanical properties. The neural network can be easily implemented in the experiments and/or in real production to resolve the uncertainty of springback amount due to the variations in material properties and friction conditions. INTRODUCTION One of the largest challenges in manufacturing is the consistency of final products. Two basic approaches have been investigated to achieve this goal. One is to use intelligent assembly methodologies to select a suitable set of parts to be assembled, that is, taking advantage of tolerance stack up. However, to our knowledge, no fundamental theory on how to efficiently select the parts has been developed and applied to various manufacturing processes. The other approach aims at each individual manufacturing process module, for example, sheet metal forming process, which is the approach and the focus of this work. As known, during unloading (tooling retreat) in the forming process, the elastic component of the stress generated during the deformation is released leading to a partial return of the deformed part toward the initial configuration. This is so-called ‘Springback’, which has been intensively studied in many recent proceedings of NUMISHEET and SAE conferences. To minimize the geometrical errors in the final shape, the reduction and the consistency of the springback are the two key issues. Springback can be reduced by a proper tooling design (die design, binder design, etc.) or by controlling the magnitude and the history of the plastic stretch imposed to the sheet by a proper binder force trajectory (process control). Karafillis and Boyce [2,3] proposed a method, “SpringForward”, for tool and binder design to obtain the desired part shape. Using finite element analysis of a forming process, the amount of springback and its associated section bending moment can be calculated and fedback to a tooling design algorithm, which provides the new tooling shape compensating the springback. The design process is repeated until reaching the desired springback amount. The “SpringForward” method provides a powerful tool for designing the tooling taking the amount of potential springback into account. The tooling geometry is often not the same as the desired part geometry. However, “SpringForward” is not designed to and therefore, not able to deal with the variations in the processes. In the aspect of the process design, Ayres [4] developed a