A Hybrid Machine Learning System for Low-Cost, Interpretable Bend, Connector, and Break Detection in OTDR Traces
Applied Sciences · 20.11.2026 · 16 (22) · 11508
DOI: 10.3390/app162211508 ↗Kurz gefasst
Fachartikel in der Zeitschrift Applied Sciences, online erschienen am 20.11.2026. Verfasst von Ilyas Kazambayev und 5 weiteren Autor:innen. Die Arbeit nennt 44 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Ministry of Trade and Integration of the Republic of Kazakhstan.
Abstract (Originalsprache)
Reliable and rapid detection of physical events along fiber-optic communication lines is essential for maintaining the availability of power-grid infrastructure. However, conventional threshold-based optical time-domain reflectometry (OTDR) analysis is hampered by weak or gradual event signatures, and modern deep-learning alternatives are often too computationally demanding for low-cost field deployment. This study presents a hybrid machine learning system for automated, interpretable detection and classification of bend, connector, and break events in OTDR traces. The system combines a dense sliding-window branch, adapted to the gradual, spatially diffuse attenuation caused by fiber bending, with a sparse candidate-generation-and-classification branch based on an ExtraTrees ensemble for the more localized connector and break events. The pipeline, together with a purpose-built annotation tool (OTDR Studio), was implemented with a Gamma Lite OTDR, a Raspberry Pi 4 for trace acquisition and transmission, and a host computer for event detection. Before splitting, all traces were assigned to heuristic measurement groups, the annotations of eight byte-identical traces were reconciled, and both branches were retrained. The system obtained a pooled (micro-averaged) event-level precision of 0.918, recall of 0.784, and F1-score of 0.846 (inverse-group-size-weighted F1-score of 0.862), on a group-level held-out test set of 81 traces with 227 reference events. Class-aware non-maximum suppression, selected on the validation set, raised the test F1-score from 0.758 (class-agnostic suppression with the same models and thresholds) to 0.846; most remaining errors are missed connector events. The system is intended as a physically grounded, interpretable tool for technicians without machine-learning expertise; standalone operation on the Raspberry Pi and field validation are future work.
Förderung
- Ministry of Trade and Integration of the Republic of Kazakhstan
Zitieren
APA
Kazambayev, I., Mekhtiyev, A., Amangeldi, M., Neshina, Y., Alkina, A., & Aimagambetova, R. (2026). A Hybrid Machine Learning System for Low-Cost, Interpretable Bend, Connector, and Break Detection in OTDR Traces. Applied Sciences, 16(22), 11508. https://doi.org/10.3390/app162211508
Vancouver
Kazambayev I, Mekhtiyev A, Amangeldi M, Neshina Y, Alkina A, Aimagambetova R. A Hybrid Machine Learning System for Low-Cost, Interpretable Bend, Connector, and Break Detection in OTDR Traces. Applied Sciences. 2026;16(22):11508. doi:10.3390/app162211508
BibTeX
@article{Kazambayev2026,
author = {Kazambayev, Ilyas and Mekhtiyev, Ali and Amangeldi, Magzhan and Neshina, Yelena and Alkina, Aliya and Aimagambetova, Raushan},
title = {A Hybrid Machine Learning System for Low-Cost, Interpretable Bend, Connector, and Break Detection in OTDR Traces},
journal = {Applied Sciences},
year = {2026},
volume = {16},
number = {22},
pages = {11508},
doi = {10.3390/app162211508}
}
EUVIMED