EUVIMEDEUVIMED entdecken
Crossref

VQ-VAE Tokenization for Anomaly Detection: A Case-Study of Induction Motor Monitoring Signals

Katarina Toroman, Xinyi Wen, Xiao Li, Benjamin Schäfer

ACM SIGEnergy Energy Informatics Review · 01.09.2026 · 6 (3) · 289-295

DOI: 10.1145/3861130.3861153 ↗

Kurz gefasst

Fachartikel in der Zeitschrift ACM SIGEnergy Energy Informatics Review, online erschienen am 11.11.2026. Verfasst von Katarina Toroman und 3 weiteren Autor:innen. Die Arbeit nennt 26 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Continuous monitoring plays an important role in predictive maintenance across industrial and energy systems. However, many modern sequence models operate most naturally on discrete symbolic inputs, whereas industrial monitoring signals are continuous and high-dimensional. It remains unclear how continuous industrial time series should be transformed into discrete representations without losing fault-relevant temporal structure. This paper studies learned tokenization as an interface between continuous monitoring signals and Transformer-based anomaly detection. In particular, we investigate Vector-Quantized Variational Autoencoder (VQ-VAE), which maps continuous signal windows into a learned, finite codebook of discrete latent tokens. We propose a two-stage pipeline in which a VQ-VAE generates token sequences and decoder-only Transformer models healthy token dynamics through next-token prediction. To isolate the effect of tokenization, the downstream Transformer architecture is kept fixed throughout the experiments while the tokenization stage is varied. Using an induction-motor benchmark as a representative energy-system case, we evaluate current and vibration signals under a normal-only training setting. The results show that learned tokenization substantially improves anomaly-detection performance over simple discretization baselines. These findings suggest that learned discrete representations are a promising interface for predictive-maintenance time series, while the most suitable tokenizer depends on signal type and representation.

Zitieren

APA

Toroman, K., Wen, X., Li, X., & Schäfer, B. (2026). VQ-VAE Tokenization for Anomaly Detection: A Case-Study of Induction Motor Monitoring Signals. ACM SIGEnergy Energy Informatics Review, 6(3), 289-295. https://doi.org/10.1145/3861130.3861153

Vancouver

Toroman K, Wen X, Li X, Schäfer B. VQ-VAE Tokenization for Anomaly Detection: A Case-Study of Induction Motor Monitoring Signals. ACM SIGEnergy Energy Informatics Review. 2026;6(3):289-295. doi:10.1145/3861130.3861153

BibTeX
@article{Toroman2026,
  author = {Toroman, Katarina and Wen, Xinyi and Li, Xiao and Schäfer, Benjamin},
  title = {VQ-VAE Tokenization for Anomaly Detection: A Case-Study of Induction Motor Monitoring Signals},
  journal = {ACM SIGEnergy Energy Informatics Review},
  year = {2026},
  volume = {6},
  number = {3},
  pages = {289--295},
  doi = {10.1145/3861130.3861153}
}
Kostenlos starten →