Supervised and unsupervised learning for classifying changes in optical time domain reflectometer traces

Global telecommunications heavily rely on optical fibers as the foundation of their network infrastructure, making it imperative for network operators to ensure their dependability. The traditional optical time domain reflectometer (OTDR) focuses on event detection, but in-service measurements can detect the interactions of distributed effects such as fiber loss, Raman amplification, stimulated Raman scattering, and channel loading. This research paper demonstrates the effectiveness of supervised and unsupervised learning models in accurately categorizing changes observed in in-service OTDR traces. Among the supervised models tested, the multilayer perceptron exhibited superior performance with a classification accuracy of 0.891 on multiple-effect data, surpassing the random forest and convolutional neural network. Clustering models were also explored, focusing on single-effect data; the best result was obtained using the Gaussian mixture model, achieving a normalized mutual information of 0.663 and an adjusted Rand index of 0.52.