EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Supervised Learning Accuracy Versus Closed-Loop Control Performance in Neural Behavioral Cloning of an Integral Sliding-Mode Controller for an Induction Motor Drive

Said Bencherifi, Azzeddine Dekhane, Mohamed Faouzi Rachedi, Habib Benbouhenni, Riyadh Bouddou, Jun-Jiat Tiang

Actuators · 06.11.2026 · 15 (11) · 592

DOI: 10.3390/act15110592 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Actuators, online erschienen am 06.11.2026. Verfasst von Said Bencherifi und 5 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Artificial intelligence (AI)-based control is gaining increasing attention in electrical-machine drives as a means of approximating nonlinear control laws through data-driven learning. Neural behavioral cloning provides a practical framework for reproducing an existing analytical controller while retaining the established control architecture. However, neural controllers are typically selected according to supervised validation loss, despite the fact that its ability to predict closed-loop control performance remains unclear. This paper systematically investigates the relationship between supervised learning accuracy and closed-loop performance through neural behavioral cloning of an integral sliding-mode speed controller for a 2.2 kW field-oriented induction motor drive. A bias-free, odd-symmetric neural network is trained to approximate the analytical control law, while the integral state, state-update law, anti-windup mechanism, actuator saturation, execution sequence, and controller lifecycle are kept identical in the analytical and neural implementations, thereby isolating the effect of the learned approximation on closed-loop behavior. The controllers are evaluated under ten benchmark conditions comprising nominal operation, six individual parameter perturbations, and a combined worst-case condition, with the sliding-surface coefficient selected equal to the PI design bandwidth parameter. Across five trained networks, whose validation mean-squared errors span a factor of 2.4, supervised validation accuracy exhibits only a weak relationship with closed-loop performance. The largest ITAE variation among the five networks within any benchmark cell is 1.60% of the corresponding cell median, while the network with the highest validation error yields the lowest ITAE in nine of the ten benchmark cells. Nevertheless, all neural controllers closely reproduce the analytical controller and achieve a 39–68% reduction in ITAE relative to a conventional proportional–integral controller, with a median reduction of 49% across the ten benchmark cells. These results indicate that, within the investigated setting, supervised validation loss is not a reliable criterion for ranking behaviorally cloned controllers according to their closed-loop performance. The findings are limited to a single induction motor drive, one teacher controller, one network architecture (H = 16), and five trained networks representing two distinct validation-error groups; broader conclusions require systematic evaluation across architectures, teacher controllers, drive conditions, and training configurations.

Zitieren

APA

Bencherifi, S., Dekhane, A., Rachedi, M. F., Benbouhenni, H., Bouddou, R., & Tiang, J. (2026). Supervised Learning Accuracy Versus Closed-Loop Control Performance in Neural Behavioral Cloning of an Integral Sliding-Mode Controller for an Induction Motor Drive. Actuators, 15(11), 592. https://doi.org/10.3390/act15110592

Vancouver

Bencherifi S, Dekhane A, Rachedi MF, Benbouhenni H, Bouddou R, Tiang J. Supervised Learning Accuracy Versus Closed-Loop Control Performance in Neural Behavioral Cloning of an Integral Sliding-Mode Controller for an Induction Motor Drive. Actuators. 2026;15(11):592. doi:10.3390/act15110592

BibTeX
@article{Bencherifi2026,
  author = {Bencherifi, Said and Dekhane, Azzeddine and Rachedi, Mohamed Faouzi and Benbouhenni, Habib and Bouddou, Riyadh and Tiang, Jun-Jiat},
  title = {Supervised Learning Accuracy Versus Closed-Loop Control Performance in Neural Behavioral Cloning of an Integral Sliding-Mode Controller for an Induction Motor Drive},
  journal = {Actuators},
  year = {2026},
  volume = {15},
  number = {11},
  pages = {592},
  doi = {10.3390/act15110592}
}
Kostenlos starten →