Metamodels based on neural networks for induction heating problems development features

Discusses the peculiarities of creating metamodels for multi-physics nonlinear processes occurring during induction heating. As an example, the development of a metamodel describing the process of induction heating when forging cylindrical billets made of 45 steel with diameters ranging from 80 to 200 mm under constant specific power mode is presented. Four stages of metamodel development are considered. In the first stage, using a numerical one-dimensional conjugate nonlinear electrothermal model, a database was created in which the input parameters defining the geometry and operating mode of the induction system and the output characterizing its operation were collected. The output parameters were divided into integral and time-distributed ones. In the second stage, the principal component method was applied to the time-distributed parameters to reduce data dimensionality. This approach enables the speed of neural network training to be increased and the number of required neurons to be reduced. In the third stage, neural networks were trained to approximate each output parameter of the database. In the final stage, the trained neural networks were integrated into a single program to demonstrate the operation of the created metamodel in real-time mode. The model based on neural networks provides good approximation of induction heating parameters in real-time, thus rendering the integration of metamodels into control systems or into packages for automatic design of induction installations highly promising.

Authors: F. V. Chmilenko, Yu. Yu. Perevalov, V. E. Parmenov, Zhang Qi, Yu. V. Shanin

Direction: Electrical Engineering

Keywords: induction heating, forging heating, metamodel, neural networks, artificial intelligence, regression, principal component analysis

---

View full article