Potential failure cause identification for optical networks using deep learning with an attention mechanism

With a focus on failure management in optical networks, we propose a potential failure cause identification scheme using an attention mechanism for optical transport network boards, leveraging actual datasets from a network operator. The attention mechanism allows the model to dynamically pay attention to only certain input information that is closely related to the target task (failure prediction), which can be effectively applied to identify the potential cause of the failure. In this paper, two typical attention mechanisms are comparatively studied to obtain the attention weights, which are additive attention and dot-product attention. A bi-directional long short-term memory network is selected as the failure prediction model due to its superior performance in time-series processing cases, which can capture bi-directional input information. Experimental results show that the average accuracy, F1 score, and false negative and false positive rates of the proposed scheme are 98.73%, 97.19%, 2.6%, and 0.91%, respectively. Moreover, based on the attention weight, it is confirmed that the highest-relevance input feature for equipment failure is the average value of input optical power, which may be caused by disconnection of the receiving port of the board or fiber cut of the adjacent link; the next most relevant feature is the minimum value of the environmental temperature, which may be caused by a broken fan or overheated chip. It is proven that the proposed scheme can not only find potential failure causes but also improve the performance of the failure prediction model, which is significant for optical networks realizing failure diagnosis and recovery.