Dynamic Range Filter and CNN–Transformer Neural Networks for OTDR Signal Event Detection

Optical fibers are susceptible to damage from external factors during prolonged use, compromising communication network performance. Optical time-domain reflectometry (OTDR) is essential for fiber-optic inspection, but conventional platforms suffer from limited noise suppression and low automation. This paper proposes a dynamic range filter tailored to OTDR traces that improves the signal-to-noise ratio by up to 22 dB while preserving reflection pulse integrity. A lightweight CNN-Transformer hybrid network (under 77 k parameters, 286 KB) is introduced for classifying reflective, non-reflective, and fiber-end events. Range-based fluctuation detection with DBSCAN clustering and specialized preprocessing are integrated to enhance reliability and precision. Tested on 51 field traces (2.5 km to 240 km, pulse widths 3–20 000 ns), the method achieves accuracies of 97.67 %, 97.22 %, and 98.03 % for the three event types, respectively, with positioning errors below 0.05 %. The proposed model achieves high detection accuracy while maintaining a compact architecture suitable for embedded deployment.