Detection, identification, and localization of faults in PONs using joint OLT and OTDR telemetry data

Passive optical networks (PONs) are a key technology for broadband delivery to residential, enterprise, and industrial internet users. Their increasing deployment in latency-sensitive environments such as manufacturing and data centers raises the importance of fast and reliable fault management. Fiber faults—whether at the telecom operator side like tapping and cutting or on the customer’s side like powered-off or unplugged optical network units (ONUs)—can disrupt service, compromise security, and violate service level agreements (SLAs). Conventional optical time domain reflectometry (OTDR) suffers from limited trace interpretability in high-split PONs, while artificial intelligence (AI)-based approaches face challenges of scalability, reproducibility, and explainability. First, we propose a heuristic algorithm that combines optical line terminal (OLT) monitoring data with OTDR traces to detect and localize faults in PONs without requiring expensive reflectors installed at the ONU or lengthy AI training and explainability issues. Then, we benchmark its classification capabilities against multiple state-of-the-art AI algorithms like convolutional neural networks (CNNs), feed forward neural networks (FFNNs), long short-term memory (LSTM), bidirectional LSTM (BiLSTM), bidirectional gated recurrent units (BiGRUs), and two different gradient boosting models. To conclude the paper, we compare our heuristic algorithm with all the others and highlight the trade-offs in accuracy, scalability, and deployment readiness. The algorithms are validated experimentally, demonstrating identification accuracy up to 90% for both deliberate and non-deliberate faults and localization error with median values below one meter. An experimental dataset is collected in this study, and it is made publicly available to support reproducibility and improve data availability.