Consequence-Aware Tail-Risk Learning for Early Warning of Rare Critical Process States
Processes · 06.11.2026 · 14 (21) · 3600
DOI: 10.3390/pr14213600 ↗Kurz gefasst
Fachartikel in der Zeitschrift Processes, online erschienen am 06.11.2026. Verfasst von Zengzheng Wang und 4 weiteren Autor:innen. Die Arbeit nennt 36 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Key Research and Development Program of China.
Abstract (Originalsprache)
Early warning of critical process states requires a future safety-boundary target and explicit control of alarm burden. Fault labels and aggregate accuracy do not directly define this objective. We propose consequence-aware tail-risk (CATR) learning, which combines mean cross-entropy (CE) with the upper tail of consequence-weighted errors. Evaluation on the Extended Tennessee Eastman Process (TEP) uses a fixed Transformer, six training compositions, five seeds, and 15, 30, and 60 min horizons. The observed critical events are reactor high-pressure crossings. Under full natural training, CATR learning increases held-out 15 min critical-window recall from 85.41% to 89.77%, reducing missed critical windows by 29.9%, with observed false-positive rates below 0.1%. The seed-by-episode 95% interval for the recall gain is [0.33, 8.79] percentage points; the average precision and macro-F1 also increase. Intervals at 30 and 60 min include zero. With the fixed 1% critical-episode training set, held-out recall is 5.20–9.54 percentage points lower than CE; size-matched natural training also yields negative mean contrasts. Separate validation experiments with resampled critical training events show sign changes at 15 and 30 min, whereas the 60 min mean contrast remains negative across the three evaluated event sets. These results support consequence-weighted tail learning for near-term warning under full natural training and identify training composition as an important condition of its performance.
Förderung
- National Key Research and Development Program of China
Zitieren
APA
Wang, Z., Zhang, X., Liu, Y., Shi, H., & Zhai, L. (2026). Consequence-Aware Tail-Risk Learning for Early Warning of Rare Critical Process States. Processes, 14(21), 3600. https://doi.org/10.3390/pr14213600
Vancouver
Wang Z, Zhang X, Liu Y, Shi H, Zhai L. Consequence-Aware Tail-Risk Learning for Early Warning of Rare Critical Process States. Processes. 2026;14(21):3600. doi:10.3390/pr14213600
BibTeX
@article{Wang2026,
author = {Wang, Zengzheng and Zhang, Xinju and Liu, Yongqiang and Shi, Hao and Zhai, Lulu},
title = {Consequence-Aware Tail-Risk Learning for Early Warning of Rare Critical Process States},
journal = {Processes},
year = {2026},
volume = {14},
number = {21},
pages = {3600},
doi = {10.3390/pr14213600}
}
EUVIMED