Self-Supervised Multi-Scale Spatiotemporal Dependency TCN-Transformer for Time Series Anomaly Detection
Algorithms · 14.11.2026 · 19 (11) · 1003
DOI: 10.3390/a19111003 ↗Kurz gefasst
Fachartikel in der Zeitschrift Algorithms, online erschienen am 14.11.2026. Verfasst von Haoran Huang und 5 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
SMST-AD is a self-supervised reconstruction framework for multivariate time-series anomaly detection that extends our previous STAM-AD model. It retains the Global/Local temporal backbone, adaptive temporal masking, synthetic anomaly insertion, and bounded anomaly loss. The extension uses parallel dilated local TCNs with independent Transformer encoders and an explicit variable-attention branch. A Pearson-correlation mask provides a fixed linear dependency prior, while anomaly-aware regularization encourages anomalous variables to retain information within the anomalous set during aggregation. Synthetic perturbations provide the training labels without requiring manually annotated faults. We evaluate SMST-AD on six public benchmarks and on constructed anomalies in real power-grid measurements, with ten independent training runs per dataset. On SMAP, SMST-AD achieves mean F1 0.9288, compared with 0.7786 for STAM-AD. In the industrial scenario, its mean F1 is 0.9318. These results demonstrate the effectiveness of SMST-AD for time-series anomaly detection.
Zitieren
APA
Huang, H., Li, B., Wang, S., Li, S., Zhang, J., & Chen, B. (2026). Self-Supervised Multi-Scale Spatiotemporal Dependency TCN-Transformer for Time Series Anomaly Detection. Algorithms, 19(11), 1003. https://doi.org/10.3390/a19111003
Vancouver
Huang H, Li B, Wang S, Li S, Zhang J, Chen B. Self-Supervised Multi-Scale Spatiotemporal Dependency TCN-Transformer for Time Series Anomaly Detection. Algorithms. 2026;19(11):1003. doi:10.3390/a19111003
BibTeX
@article{Huang2026,
author = {Huang, Haoran and Li, Baowei and Wang, Shaopeng and Li, Shuyue and Zhang, Jianjun and Chen, Bolin},
title = {Self-Supervised Multi-Scale Spatiotemporal Dependency TCN-Transformer for Time Series Anomaly Detection},
journal = {Algorithms},
year = {2026},
volume = {19},
number = {11},
pages = {1003},
doi = {10.3390/a19111003}
}
EUVIMED