Dirty-data-based alarm prediction in self-optimizing large-scale optical networks

Machine-learning-based solutions are showing promising results for several critical issues in large-scale optical networks. Alarm (caused by failure, disaster, etc.) prediction is an important use-case, where machine learning can assist in predicting events, ahead of time. Accurate prediction enables network administrators to undertake preventive measures. For such alarm prediction applications, high-quality data sets for training and testing are crucial. However, the collected performance and alarm data from large-scale optical networks are often dirty, i.e., these data are incomplete, inconsistent, and lack certain behaviors or trends. Such data are likely to contain several errors, when collected from old-fashioned optical equipment, in particular. Even after appropriate data preprocessing, feature distribution can be extremely unbalanced, limiting the performance of machine learning algorithms. This paper demonstrates a Dirty-data-based Alarm Prediction (DAP) method for Self-Optimizing Optical Networks (SOONs). Experimental results on a commercial large-scale field topology with 274 nodes and 487 links demonstrate that the proposed DAP method can achieve high accuracy for different types of alarms.