Speaker
Description
Open-die forging of ingots is one of the main hot metalworking operations in the metal industry. The operation improves the quality of the material by void removal and grain structure refinement, especially in the cogging process. To access the material quality, the internal strain evolution and final distribution are important. Previous work has focused on the use of empirical knowledge, FEM simulations and more recently neural network approaches. This work focuses on the neural network approach, with development of a 3D, fully data-driven artificial neural newtwork (ANN).
A quarter symmetric 3D FEM was used as a basis for generating training and validation datasets for the creation of ANNs. Two fully data-driven ANNs were created to predict the geometry change and equivalent strain evolution during the cogging process of rectangular cross-sections. A unique approach of individual node prediction was used to minimize calculation times and generate results instantaneously.
Preliminary results show that the geometric and strain models perform well, with less than 2 % (RMSE) prediction errors for nodes in the volume, while surface nodes experience larger error in the range of up to 10 % at nodes with high deformation.
The method developed appears promising for the prediction of geometry and equivalent strain and future work will focus on new geometries and variation of additional parameters and forging operations.
| Speaker Company/University | Luleå Univeristy of Technology |
|---|