Neural-network feedback control of an extrusion

This work is concerned with the feedback control of microstructure during one of the simplest metal forming operations: round-to-round extrusion. Physically based semiempirical models of the microstructural dynamics are available, but they require flow variables such as strain, strain rate, and temperature as inputs. Direct measurement of these quantities inside the deforming material is not feasible, so such models alone do not define a feedback controller. In the study presented, the mapping from the temperature of the material flowing through the die to the ram load is estimated via finite-element simulation. The ram load can be measured, and so this mapping, composed with the microstructural model, does close the loop, but the simulation is far too slow for real-time implementation. This problem is addressed by training an artificial neural network to represent the simulation output. This approach is demonstrated on the simulated extrusion of a plain carbon steel rod.