Machine learning approaches for classical optical time-domain reflectometer: A review

The optical time-domain reflectometer (OTDR) is a key instrument for characterizing optical fiber links and locating faults. However, conventional OTDR trace analysis still depends heavily on manual interpretation and signal-processing methods, which often struggle under noisy, low signal-to-noise ratio, and deployment-dependent conditions. Machine learning (ML) has therefore emerged as a promising approach for automating OTDR analysis and improving the detection, characterization, and localization of fiber events. This paper presents a focused review of ML methods for classical OTDR, where classical OTDR refers to intensity-based OTDR used for link characterization and fault localization rather than phase-sensitive distributed sensing. The review covers OTDR fundamentals, the limitations of traditional analysis, and the main OTDR tasks addressed by ML, including data preparation, event detection, anomaly handling, localization, branch assignment, and operational monitoring. It also summarizes the main model families reported in the literature, namely classical, deep, and hybrid approaches, together with commonly used datasets and evaluation metrics. Overall, the reviewed studies show that classical ML remains useful when feature engineering is effective and data are limited, whereas deep and hybrid models tend to perform better in noisy, multitask, and passive optical network scenarios. The literature also highlights the growing importance of data robustness, domain adaptation, and physics-aware preprocessing for reliable real-world OTDR automation.