Interpolation of Forging preform shapes using neural networks

Abstract An important aspect of the forging process is the design of preforms (or blockers) to achieve adequate metal distribution. In many cases determination of the preform configuration is a difficult task and art requiring skills acquired over many years. Currently available support is from skilled craftsmen, Finite Element (FE) simulation for axisymmetric components and Expert Systems. Each of these provides Inadequate support. The proposed research expects to establish a new technique — the interpolation of preform shapes for a component from manufacturing information for the family to which this component belongs. The technique will be proven by referring to the processing requirements of a family of plane-strain symmetrical H - spahed products. The research aims to establish an unified approach to use existing knowledge about the preform design, FEM simulation results and physical modelling experimental results to train a backpropagation feed forward neural network. The trained network is expected to interpolate within the component family to predict preform shapes. Exact dimensions for the preform can be determined by analytical approach or expert systems. This would reduce the involvement of time consuming FEM analysis and physical modelling for the design.