Data-driven discovery of the heat equation in an induction machine via sparse regression

Saeideh Khatiry Goharoodi, Pieter Nguyen Phuc, Luc Dupré, Guillaume CrevecoeurIcit 2019 · DOI

Introduction

In engineering, physical laws like the heat equation are typically used as a starting point to determine specific parameters such as thermal conductivity. However, "automated discovery" seeks to invert this process, allowing experimental data to reveal the underlying mathematical laws. While symbolic regression has been used for this purpose, it is often computationally prohibitive for large datasets. Sparse regression techniques, which select the simplest model from a library of candidates, offer a more efficient alternative. This study investigates whether these algorithms can successfully identify the partial differential equations (PDEs) governing a complex, noisy industrial system: a 5.5 kW induction motor.

Experimental Setup and Data Collection

The research utilized primary experimental data from a Totally-Enclosed-Fan-Cooled (TEFC) 5.5 kW induction machine. Four PT1000 RTD sensors were embedded at various axial positions within the stator windings to capture temperature gradients.

Experimental setup showing the placement of four RTD sensors (T1-T4) along the axial direction of the stator windings.

Two primary experimental scenarios were conducted:

  1. Scenario A (Active Cooling): The motor and its cooling fan operated for approximately 3.5 hours before shutdown.
  2. Scenario B (Natural Cooling): The motor operated without the fan for approximately 6.6 hours.

The resulting datasets comprised 2,500 and 4,500 samples, respectively, providing a robust basis for testing the discovery algorithm against real-world noise and thermal dynamics.

Measured temperature profiles over time for the four sensors during the active cooling scenario (Scenario A).

Methodology: Sparse Identification of Nonlinear Dynamics

The authors adapted the SINDy (Sparse Identification of Nonlinear Dynamics) framework to handle experimental constraints. The methodology involves three critical phases:

  1. Numerical Differentiation: Temperature ($T$) was differentiated with respect to time ($t$) and axial position ($z$). A pseudo second-order accurate scheme was employed to account for non-uniform sensor spacing.
  2. Library Construction: A candidate matrix ($S$) was built, containing potential terms for the PDE, including linear and nonlinear combinations of $T$ and its derivatives (e.g., $T, T^2, T_z, T_{zz}$).
  3. Sparse Group Lasso: To overcome the high correlation between candidate terms—a common issue in experimental data—the researchers used Sparse Group Lasso. This technique groups related terms and effectively "zeros out" irrelevant ones, isolating the physically meaningful components of the equation despite significant noise.

Results and Discussion

The algorithm successfully "rediscovered" the 1D heat equation from the raw sensor data: $$T_t = c_1 T_{zz} + c_2$$

The discovered coefficient for thermal diffusivity ($c_1$) was approximately $8.16 \times 10^{-5}$ for Scenario A and $7.86 \times 10^{-5}$ for Scenario B. These values show acceptable agreement with the theoretical value of $7.18 \times 10^{-5}$ derived from the material properties of copper and resin.

Key Insights

  • Robustness to Real-World Noise: Unlike many studies relying on synthetic data, this work demonstrates that sparse regression can function on a vibrating, electromagnetic industrial machine.
  • Invariance to Boundary Conditions: The algorithm identified the same underlying PDE regardless of the cooling state (fan on vs. off), indicating it captured the fundamental physical law rather than over-fitting to specific experimental conditions.
  • Material Property Estimation: The coefficients $c_1$ and $c_2$ correspond to physical constants like thermal diffusivity and internal heat generation, suggesting that sparse regression can be used for in-situ material characterization.

Conclusion and Future Work

This study confirms that data-driven methods can identify fundamental physical equations in complex industrial hardware. While successful in a 1D axial model, the current approach is limited by sparse spatial sampling. Future research will focus on expanding the sensor array to capture 2D and 3D heat flow (radial and circumferential) and applying these techniques to more complex, non-linear physical phenomena within industrial machinery.

Summary updated

BibTeX
@inproceedings{goharoodi2019data,
  author = {Saeideh Khatiry Goharoodi and Pieter Nguyen Phuc and Luc Dupré and Guillaume Crevecoeur},
  title = {{Data-driven discovery of the heat equation in an induction machine via sparse regression}},
  booktitle = {Icit},
  year = {2019},
  doi = {10.1109/icit.2019.8754983},
}
Figure