Publication

MDO & PHM Lab.

Journal

Domain subset ensemble adversarial network and modified supervised adversarial autoencoder for enhancing domain generalization of induction motor fault diagnosis
Year
2026
Author
Sukeun Hong, Jongsoo Lee
Journal
Journal of Supercomputing
Vol.
82
Issue Date
2026.01
Industrial induction motor fault diagnosis requires robust domain generalization (DG) models capable of real-time operation under various operating conditions. Existing DG methods commonly adopt a design in which a single model learns domain-invariant representations from all source domains. However, for motors operating across a wide range of speeds, fault characteristics that remain stable at low speeds differ fundamentally from those at high speeds. The challenge of capturing diverse fault signatures has motivated an ensemble approach that combines a domain subset ensemble adversarial network (DSEAN) with a modified supervised adversarial autoencoder (mSAAE). The DSEAN systematically trains complementary models from adjacent-domain subsets, where each model learns localized invariant features optimized for specific operating ranges. Each model was trained independently, enabling parallel training and ensemble diversity. The mSAAE complements this via continuous domain interpolation using continuous rotational speed labels, bridging gaps in discrete source distributions. This combination addresses both discrete sampling limitations through data augmentation and representational diversity through ensemble learning, providing a comprehensive coverage of continuous operating ranges. Validation on a bearing benchmark dataset demonstrated an average accuracy of 97.42%, outperforming nine state-of-the-art methods by 0.59%. This methodology is particularly effective for interpolation scenarios, although challenges remain in extrapolation tasks. The modular architecture enables incremental updates, making it suitable for high-performance computing environments where real-time fault diagnosis and computational scalability are essential.