Knowledge-Based Collective Self-learning for Alarm Prediction in Real Multi-Domain Autonomous Optical Networks

In this paper, a collective self-learning method based on knowledge sharing is proposed to predict alarms in multi-domain autonomous optical networks. The well-considered architecture is rendered, together with various alternatives for combining machine learning (ML) knowledge. The proposed method has been tested in the commercial large-scale multidomain network with 274 nodes and 487 links. Experimental results show that it can achieve high accuracy for alarm prediction. In addition, it can achieve similar performance with much better flexibility than a collective scheme based on training data sharing as well as more superior accuracy and robustness than an individual ML model.