Optical fiber cables deployed in the last 30 years have been considered to be long-lived components due to the generally high level of reliability of glass and the robustness of cable construction. Like all other components however, optical cables can change with age as a result of their construction or deployment characteristics as well as operational and environmental stress factors. In this paper, we apply a time-series decomposition method on span loss data collected during 12 months in four bidirectional spans of a production network in order to detect long-term degradation of the fiber plant. After extracting the trend component, a Mann-Kendall test is applied to the span loss curves and a Sen's slope estimator test is used for determining the magnitude of span loss change. The method allows detecting a 1.3 % increase in span loss over the 12-month period of observation in one of the fiber spans. This trend is confirmed using a linear regression model computed according to Theil Sen method applied to additional one-year data. The proposed method can allow the early detection of fiber plant degradation and the proactive planning of fiber replacement.

In a context of ever-increasing traffic, a degradation of the optical layer could affect client demands, in particular the quality of service provided by telecommunications operators. Thus, the rapid detection and prediction of performance degradations occurring in the optical lightpath could help to minimize errors in the network. This paper proposes a failure detection model, equivalent to a performance degradation detection model, but based on machine learning (ML) techniques, namely, the interquartile range (IQR) and the support vector machine (SVM) methods. Note that this model is built from performance metrics monitored on real optical lightpaths. In addition, our model can both label the anomalies to be defined on the data and capture the features that will be used. Feature engineering is explored using three ML techniques, namely the Boruta algorithm, the Random Forest classifier and the recursive feature elimination (RFE), to select the most useful features for the implementation of the model. Tested on monitored performance metrics, the validation phase shows that the model using the RFE method gives us the best results with an F1-score and a recall of 99.51% and 100%, respectively. These results prove the model's ability to detect in advance the degradation of the performance of the network.