Modular neural networks for quality of transmission prediction in low-margin optical networks

It is estimated that the emergence of digital businesses, and growing dynamic traffic demands for Internet applications, including cloud computing and the Internet of Things, will impact the traffic at unprecedented rates. Thus, to cost-efficiently accommodate these challenging requirements, network operators are motivated to maximize the achieved capacity over deployed links, and design more efficient networks capable of handling a wide range of applications, as well as operate more closely to optimality. On the other hand, as the network elements age, nonlinear impairments increases (which can be translated to increasing the network load) and hence, the system margin and maximum achievable rate decrease. Therefore, in order to increase the capacity of optical networks and extend their life time, an accurate physical model is required to estimate the quality of signal and quantify the margin of lightpaths. In this regard, a machine learning (ML) method is proposed based on the modular neural networks to account the cross channel nonlinear effects in estimating noise power and OSNR of lightpaths. For this, the dependence of the contributed nonlinear component of noise power on the transmitted powers as well as the modulation formats of all channels are considered in this data-model. Indeed, this ML-based model provides a useful tool for per-lightpath power management, as a fundamental element for reducing margins. Results of evaluating various proposed modular models show that, in general, all modular models are able to estimate OSNR of lightpaths, with average accuracy of more than 96.6%. However, the proposed $$ MM_{4} $$ M M 4 modular model is the model presenting better generalization, being able to correctly estimate OSNR of lightpaths, with average accuracy of 99.2%. Also, $$ MM_{3} $$ M M 3 model performs better for partially loaded network scenarios (99.3%).