Soft Failure Identification for Long-haul Optical Communication Systems Based on One-dimensional Convolutional Neural Network

Abstract

With the advance of elastic optical networks, optical communication systems are becoming more flexible and dynamic. In this scenario, soft failures are more likely to occur due to various link impairments. If these soft failures are not handled properly and timely, service disruption may occur. Identifying the cause of soft failure is a key step to restore the degraded links. However, it is difficult for traditional methods to accomplish this task. Fortunately, powerful machine learning (ML) algorithms provide a promising path to address this problem. In this article, a novel two-stage soft failure identification scheme based on a convolutional neural network (CNN) and receiver DSP is proposed. The input of the CNN is the power spectrum density (PSD) extracted from a coherent receiver, and the output contains the identified cause of soft failures together with their probabilities. Extensive simulations are performed to validate the proposed method. Four types of soft failure causes are explored including the offset of optical filter's center frequency (FS), the tightening of optical filter's 3-dB bandwidth (FT), SNR degradation due to the increased amplified spontaneous emission (ASE) noise, and the Kerr nonlinear effect. When only one soft failure cause exists, excellent accuracy is achieved. When multiple soft failure causes exist, the probabilities of these causes provided by the CNN are used to gain insight into their influences on the system. Finally, we investigate the interpretation of the CNN and a reasonable interpretation is given and discussed.