Fiber Optic Fault Detection Based on Optical Time Domain Reflectometer Data Using Random Forest Algorithm

Faults in fiber optic networks, such as fiber cuts, bending, and bad splices, can significantly degrade service quality if not detected quickly and accurately. Manual visual interpretation of Optical Time Domain Reflectometer (OTDR) traces by technicians is slow and prone to subjective error, particularly in large-scale networks. This study aims to build an automatic classification system to detect eight fiber optic network conditions (normal, fiber tapping, bad splice, bending, dirty connector, fiber cut, PC connector, and reflector) based on OTDR data using the Random Forest algorithm. The features used include Signal-to-Noise Ratio (SNR), a thirty-point trace sequence (P1 to P30), event position, reflectance, and loss. We split the dataset into training and testing sets using an 80:20 ratio. We tuned hyperparameters via GridSearchCV. Test results show that the Random Forest model with 200 decision trees and an entropy splitting criterion achieves 99.99 percent accuracy, a 99.99 percent macro F1-score, and a perfect ROC-AUC (1.0000) across all eight classes. The loss and Reflectance features proved to be the most influential predictors in the classification. Additional validation on 301 real-world OTDR field measurement files revealed a domain shift relative to the synthetic training data, indicating that the model’s generalization to field conditions without recalibration remains a limitation of this study. The resulting system is implemented as an interactive interface on Google Colab that is publicly accessible, allowing users to upload new OTDR data and obtain classification results along with the model’s confidence level directly.