Deep-learning-based failure prediction with data augmentation in optical transport networks

Abstract

Failures in optical transport networks usually result in lots of services being interrupted and a huge economic loss. If the failures can be predicted in advance, some actions can be conducted to avoid the above adverse consequences. Deep learning is a good technology of artificial intelligence, which can be used in many scenarios to replace humans’ activities. Event prediction is a typical scenario, where deep learning can be used based on a large dataset. Therefore, deep learning can be used in optical transport networks for failure prediction. However, dataset construction is an important problem for deep learning in optical transport networks, because there may be not enough data in reality. This paper proposes a deep-learning-based failure prediction (DLFP) algorithm that constructs available dataset based on data-augmentation for data training. DLFP algorithm is composed of alarm compression, data augmentation, and fully-connected back-propagation neural network (FCNN) algorithm. Besides, a benchmark algorithm (BA) without data augmentation is introduced. A training model is constructed based on massive real performance data and related alarm data within one month, which are collected from national backbone synchronous digital hierarchy (SDH) network with 274 nodes and 487 links in China. Then the training model is used with test dataset to verify the performance in terms of prediction accuracy. Evaluation results show that the proposed algorithm is able to reach better performance for failure prediction compared with the benchmark without data augmentation.