Hard-Landing Risk in Touchdown-Load Physics: A Multiple Instance Learning Framework for Interpretable Cause Attribution From QAR Data
IEEE Access · 2026 · 14 · 106943-106954
DOI: 10.1109/access.2026.3706790 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Mingming Yang und 4 weiteren Autor:innen. Die Arbeit nennt 34 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Key Research and Development Program of China, Postgraduate Research and Practice Innovation Program of Jiangsu Province, Scientific Research Program of Universities of Inner Mongolia Autonomous Region.
Abstract (Originalsprache)
Hard landings impose vertical loads beyond design limits and are a primary precursor to runway excursions, structural fatigue, and unscheduled maintenance. Existing per-flight attribution on Quick Access Recorder (QAR) data reports contributing parameters or deviation windows separately, with their link to landing physics often limited to empirical association. This paper proposes a physics-informed framework that jointly assesses hard-landing risk and identifies its cause through three modules. A mechanism module turns the closed-form touchdown vertical-load equations into a 12-parameter kinematic reference set and enforces it on the model’s feature gate, constraining the network to the physical drivers of touchdown load. An attention module, a Gated Attention Unit Transformer with max-pooling multiple instance learning, produces a bag-level prediction and a per-instance risk sequence. An attribution module identifies the dominant deviating parameter and its time window by deviation analysis against a normal-cohort envelope. On 2,075 Airbus A320 flights, four independent measurement channels agree on the kinematic set in 82.4% of dominant-parameter selections, with the gate regularizer concentrating 94.1% of the feature-gate mass on the same set. The proposed encoder reaches the highest per-instance score concentration, with a coefficient of variation 3.4 to 30 times that of diffuse-attention alternatives, and cascading the attribution alarm after the classifier raises recall from 0.964 to 0.994. For each flagged hard landing, the framework reports the dominant parameter, its time window, and the contributing approach sub-phase, all traceable to landing physics, supporting parameter-level FOQA review.
Förderung
- National Key Research and Development Program of China
- Postgraduate Research and Practice Innovation Program of Jiangsu Province
- Scientific Research Program of Universities of Inner Mongolia Autonomous Region
- Sichuan Science and Technology Program
- National Natural Science Foundation of China
Zitieren
APA
Yang, M., Peng, S., Ren, X., Peng, X., & Shao, Q. (2026). Hard-Landing Risk in Touchdown-Load Physics: A Multiple Instance Learning Framework for Interpretable Cause Attribution From QAR Data. IEEE Access, 14, 106943-106954. https://doi.org/10.1109/access.2026.3706790
Vancouver
Yang M, Peng S, Ren X, Peng X, Shao Q. Hard-Landing Risk in Touchdown-Load Physics: A Multiple Instance Learning Framework for Interpretable Cause Attribution From QAR Data. IEEE Access. 2026;14:106943-106954. doi:10.1109/access.2026.3706790
BibTeX
@article{Yang2026,
author = {Yang, Mingming and Peng, Shiyi and Ren, Xiaosong and Peng, Xiaolin and Shao, Quan},
title = {Hard-Landing Risk in Touchdown-Load Physics: A Multiple Instance Learning Framework for Interpretable Cause Attribution From QAR Data},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {106943--106954},
doi = {10.1109/access.2026.3706790}
}
EUVIMED