Estimation and Reduction of Inter−Channel Interference and Nonlinear Effects in Nyquist−WDM systems using Machine Learning

Abstract

ABSTRACT : The increment of data traffic demand by end users will require to increase the capacity of next generation optical networks, both, long−haul and optical access networks (OANs). The use of high-level modulation formats, wavelength division multiplexing (WDM) and digital coherent receivers, will enable the increase of the OANs capacity. On the other hand, for long-haul networks, the paradigm called Elastic Optical Networks (EONs) will be essential to increase the capacity, thanks to the dynamic use of networks resources. For instance, a dynamic use of the optical spectrum will enable the gridless networks which jointly with Nyquist pulses that present rectangular spectral shapes, will improve the Spectral Efficiency (SE) and thus, the capacity of the networks. Nevertheless, nonlinear phase noise, Kerr effects and inter-channel interference (ICI), are some of the challenges that future OANs and gridless networks would face. The nonlinear effects are present when high launch power (needed for high order modulation formats) is induced into the optical fiber. These effects distort the transmitted symbols generating errors after demodulation and, thus, increasing bit error rate (BER). On the other hand, the ICI effects increase the BER due to the interaction of the optical channels when the spectral spacing is reduced up to the baud rate. Machine Learning (ML)−based techniques have shown improvement of traditional monitoring and mitigation of different effects in optical communications. ML is regarded as one of the most promising methodological approaches to perform network-data analysis and has shown relevant early contributions in optical communications, allowing channel estimation, mitigation of nonlinear effects and foresees to be a powerful tool into the optical performance monitoring (OPM) field. Thus, the goal of this work is to develop techniques using ML algorithms for i) estimation of spectral spacing in gridless Nyquist-WDM systems and ii) mitigation of nonlinear and ICI effects. For the former, two spectral overlapping estimation tools are proposed, both use information obtained from fuzzy clustering algorithms applied to 10k received symbols frames. The first one uses the membership degrees of the partition matrix resulted by an unsupervised learning algorithm to construct different features which are called "counting vectors". The second uses clustering validation indexes after the unsupervised learning algorithms are applied. It is tested two unsupervised learning algorithms that use fuzzy clustering: Fuzzy cMeans (FCM) and Gustafson-Kessel Means (GKM). On the other hand, in order to reduce nonlinear and ICI effects, three supervised learning algorithms: Support Vector Machine (SVM), Artificial Neural Networks (ANN) and k-Nearest Neighbors (KNN) are proposed to carry out nonsymmetrical demodulation and thus, minimize the BER. Experimental results show that it is possible to determinate if any constellation diagram is affected by spectral overlapping using any of the proposed methods. Accuracies up to 91% are achieved in binary classification knowing a priori the OSNR value in the case of 32 GBd with fiber transmission. Moreover, when the OSNR is not known by the classifier, accuracies up to ~80% are obtained using the 16 GBd data. Besides, classification of different scenarios as high or low spectral overlapping jointly with detection of close channels or single channel is obtained also. Whereby, results show that it is possible to mitigate nonlinear effects as well as ICI effects achieving gains in terms of the OSNR. Additionally, the BER reduction is obtained at low training lengths for all the three proposed methods. Simulation results show the effectiveness of the methods by presenting gains up to ~0.5 dB in terms of the OSNR at a FEC limit of log10(BER) = −2.4. Besides, at high launch power, traditional demodulation method reaches BER values of log10(BER) = −2 while the proposed methods obtain ~log10(BER) = −2.6. Moreover, the experimental results show mitigation of ICI effects achieving gains up to ~4 dB in terms of OSNR at spectral overlapping of 18%. Consequently, it is demonstrated that the use of ML algorithms will be useful in future OANs and gridless networks by allowing mitigation of signal impairments and could be helpful to control lasers frequencies by estimating the spectral overlapping among channels.