Thermal Model Calibration of a Squirrel-Cage Induction Machine

TL;DR

This paper addresses accurate thermal modeling of induction machines by calibrating a 2D finite-element model through an inverse thermal field problem to implicitly incorporate 3D effects. The forward problem solves a static heat conduction PDE with temperature-dependent conductivities using a 2D FE discretization, while homogenization and parameter calibration compensate for neglected axial conduction and cooling. The method is validated on two academic test cases with synthetic data and then applied to a real induction machine, achieving good agreement with independent measurements and demonstrating reliable hot-spot prediction. The work enables efficient, data-driven thermally informed design and operation by reducing reliance on large safety margins while accounting for three-dimensional effects through parameter calibration.

Abstract

Accurate and efficient thermal simulations of induction machines are indispensable for detecting thermal hot spots and hence avoiding potential material failure in an early design stage. A goal is the better utilization of the machines with reduced safety margins due to a better knowledge of the critical conditions. In this work, the parameters of a two-dimensional induction machine model are calibrated according to evidence from measurements, by solving an inverse field problem. The set of parameters comprise material parameters as well as parameters that model three-dimensional effects. This allows a consideration of physical effects without explicit knowledge of its quantities. First, the accuracy of the approach is studied using an academic example in combination with synthetic data. Afterwards, it is successfully applied to a realistic induction machine model.

Figures (8)

• Figure 1: Inverse parameter estimation process. Beginning with an initial guess $\theta_0$, the estimated parameters $\theta_\text{e}$ are obtained in an iterative process.
• Figure 2: Geometry of example 1. The model consists of a domain with thermal conductivity $\lambda_1$ and a domain with generated heat $g$ and thermal conductivity $\lambda_2$.
• Figure 3: Geometry of example 2 with discrete sensors, visualized by dotted circles. In addition to example 1, a cylindrical layer with thermal conductivity $\lambda_3$ is added.
• Figure 4: Relative error of example 1 over number of independently perturbed simulation with multiple seeds.
• Figure 5: Relative error of example 2 over number of independently perturbed simulations.