ML-driven classification of link components in passive optical networks
Telecom operators deploy and operate large amounts of passive optical networks (PONs) delivering high-speed broadband internet to homes and small businesses. The maintenance and high-reliability requirements for such networks is a challenging task, helped by specialized fiber monitoring equipment such as optical time domainreflectometer (OTDR). This thesis is focused on analyzing and interpreting OTDR traces using machine learning (ML) techniques. An OTDR trace data set of varying PON
... res is collected in a laboratory setup using commercial OTDR equipment. Two ML approaches for event identification and localization in OTDR traces are compared. The results show that a two-stage approach using a deterministic event localization algorithm and an ML based identification model comprehensively outperforms the one-stage approach of using a single ML model to localize and identify events. An accuracy of 98% is achieved for the important task of identifying optical network units (ONUs) while the accuracy over all types of events lies at ~90% . The gained information can be used to find problems and failures to reduce costs and increase work efficiency. Current state-of-the-art techniques mostly only use deterministic algorithms to detect and classify events happening in PONs. This approach is not ideal for PONs because with slope calculation a classification cannot be made. Additionally, events can overlap in PONs and various levels of attenuation and noise result in different looking traces, making analysis difficult. Convolutional neural networks, however, can extract common characteristic to correctly classify each event. In this work, a deterministic algorithm is used to propose a position for a potential event which is to be classified by the convolutional neural network.