Hybrid Deep Learning Model for Resolving Overlapping Events in OTDR Dead Zones

Conventional optical time-domain reflectometry (OTDR) suffers from event and attenuation dead zones when strong Fresnel reflections saturate the receiver, obscuring closely spaced events and degrading localization accuracy. High-performance OTDRs mitigate these issues by using ultrashort pulses, high-bandwidth detectors, and low-noise front ends, but at the expense of increased cost and calibration complexity. This work introduces a hybrid deep learning framework that enhances the sensing capabilities of a low-cost OTDR without modifying its hardware. An experimental dataset of 2150 traces was collected from polymer optical fibers subjected to controlled microbending loads at variable separation distances. The proposed model fuses waveform- and feature-based representations through convolutional, bidirectional long short-term memory, and attention encoders to resolve overlapping events within OTDR dead zones. It achieves 100% event-count classification and subdecimeter localization accuracy (mean absolute error < 0.09 m), providing measurable performance gains relative to conventional signal interpretation. These results demonstrate that data-driven OTDR evaluation can reduce ambiguity in dead zones and extend the practical functionality of low-cost distributed optical sensors, thereby supporting the development of intelligent cost-effective monitoring systems.