Speaker
Description
Modeling of induction hardening is a complicated task due to involvement of multiple physical processes. One has to solve the highly nonlinear coupled electromagnetic, thermal, mechanical and metallurgical problems with proper boundary conditions. Finite element methods (FEM) are a good choice for a physics-based description of complex systems such as Induction hardening. However, the full details of the interactions and process variables are crucial for an accurate modeling, but many of the required details are often not available. This imposes a true challenge to optimize induction hardening processes.
Black-box models like artificial neural networks (ANNs) are the tool of choice when it comes to finding nonlinear relations between process parameters and material response. However, the training of ANNs is costly and they eventually might not be applicable efficiently in practice. To circumvent these limitations the ANNs are integrated with available knowledge of the system. The knowledge is based on the known physical and conservation laws governing the process.
In our work we construct and test dynamic hybrid models (HMs) to simulate the temperature evolution of a layer located under the surface of an inductively heated cylindrical sample. The HM has a serial structure in which the ANN serves as an estimator of a non-observable process parameter (heat source term) embedded in the physical model. The unknown parameter estimated from the ANN is temperature dependent and depends nonlinearly on other process parameters such as operating power and material data which is identified and approximated from the training data examples. The physical equation is then integrated to estimate the temperature in the heating time for a given location in the sample. We train and optimize different HM variants with experimental data acquired on our induction heating test rig. We analyse the quality and ability of the models to explain the data.
| Speaker Country | Austria |
|---|