Studies On Key Technologies Of Optical Performance Monitoring And Signal Processing For Intelligent Cognitive Optical Networks

Recently,with the development of emerging network services,complex advanced modulation,dynamic wavelength routing,flexible spectrum grid and mixed transmission technologies,the dynamicity,complexity and heterogeneity of optical networks have been improved increasingly,which has posed greater challenges to the network management and controlling.In the traditional optical networks,the adaptivity of transmission systems is limited and normalized monitoring data are not sufficient for the network management,which also lacks the cross-layer cognition and intelligent feedback mechanism between the physical transmission systems and the network management and controlling.To address those problems,with the support of the artificial intelligent(AI)and software defined networking techniques,the cognitive optical network(CON)has the potential to be equipped with the comprehensive network state monitoring ability,intelligent network data analysis functions and adaptive network element controlling advantages,which has already been one of the research hotspots in current optical networks.This dissertation focuses on the optical performance monitoring and signal processing technologies in intelligent cognitive optical network.Here,the AI-driven intelligent cognitive optical network architecture,adaptive impairment monitoring and compensation for transmission systems,data augmentation for intelligent cognitive network database,and network resource management based on the physical infrastructure cognition and the network layer feedback mechanism are comprehensively exploed and reaserched.Several technical innovation schemes are proposed.The research highlights of this paper are listed as follows:Firstly,an AI-driven intelligent cognitive optical network architecture is specially designed based on the key idea called "cognition-learning-action" cognitive loop in the CON,where the data source,monitoring approaches,data storage and data representation are normalized in the network state cognition.Moreover,diverse AI-driven applications are introduced in the network management and controlling,including the optical performance monitoring,physical impairment compensation,network element controlling,quality of transmission estimation,network resource allocation,network traffic prediction and network fault management.What's more,the principal and implementation schemes of different software-defined network elements are concluded.The closed-loop "cognition-learning-action" functions are realized in the proposed AI-driven network infrastructure,which provides the network infrastructure fundamental for the intelligent CON.Secondly,focusing on the problems of current physical impairment compensation algorithms in terms of pool intelligent learning ability,transmission link information dependence and the limited adaptivity,two machine-learning-based adaptive system impairment monitoring and compensation approaches are developed.For the optical fiber transmission system,a deep neural network(DNN)-based adaptive chromatic dispersion(CD)monitoring scheme is proposed.The simulation results show that the average monitoring error is near 20ps/nm within 2000ps/nm dynamic monitoring range.Compared with the classic scheme based on the CD scanning and frequency domain equalization,the computation complexity of the proposed adaptive CD monitoring scheme is obviously decreased,where the number of the adders,multipliers and dividers respectively decreases by 98.6%,98.8%and 64.4%.Moreover,for the free space optical transmission system,a joint atmospheric turbulence monitoring and adaptive demodulation scheme based on the convolutional neural network(CNN)is proposed for the orbital angular momentum key shifting(OAM-SK)communication system.The adaptive demodulation error of the proposed scheme is near 0.86%in the 8-OAM-SK system.Compared with the traditional scheme,the modulation accuracy is improved by 19.2%.Meanwhile,the atmospheric turbulence monitoring is firstly realized by the CNN and the monitoring accuracy is 95.2%for 6 kinds of typical atmospheric turbulence channels.Thirdly,aiming at the problems of the insufficient normalized network traffic data for AI-driven applications and there is serious data imbalance in network fault management data,an adaptive sequential data augmentation algorithm based on deep learning and a network fault data balance algorithm using the generative adversarial network(GAN)are proposed.Experimental results show that the proposed traffic data augmentation algorithm is effective for 6 kinds of typical traffic scenarios.Compared with the statistical parameters of the corresponding practical traffic,the deviations about the mean,variance and Hurst exponent of augmented traffic data from the proposed algorithm is respectively 0.7%,1.3%and 7.0%,which is obviously less than those of the classical statistical parameter configuration(SPC).Moreover,the experimental results show that the false positive rate can be decreased from 24.7%to 3.8%when the proposed GAN-based network fault balance algorithm is adopted in the classical network fault recognition scheme based on support vector machine(SVM).For the network fault recognition schemes using different AI algorithms including SVM,k-th nearest neighbor(KNN),decision tree(DT),random forest(RF)and gradient boosting decision tree(GBDT),the GBDT-based network fault recognition approach combined with the proposed GAN-based data balance algorithm exceeds obviously than other algorithms in terms of the false positive rate,accuracy and recall rate comprehensively,which can effectively decrease the impacts of the data imbalance in the network fault recognition.Fourthly,focusing on the problems of optical network resource management and controlling methods in terms of the manual intervention dependence,deficiency in network feedback and bad dynamic modelling ability,a network resource management mechanism based on the physical infrastructure cognition and network feedback is proposed,where the digital twin technique enabled by the deep reinforcement learning(DRL)is introduced to improve the dynamic modelling and intelligent controlling capabilities of the programmable optical transceiver(POT)in which the modulation format,symbol rate and feedforward error correction(FEC)coding can be automatically adjusted on demands.Compared with the classic POTs based on the maximum capability(MaxCap),the proposed DRL-based POT is capable of saving 19.4%spectrum resource while the similar network delay performance can be obtained.Benefited from the double neural networks and the feedback controlling mechanism in the double-agent DRL,the proposed DRL-POT can model the POT dynamically according to the time-varied traffic load and quality of transmission and further provide optimal POT action with maximum utilization to control the physical POT to guarantee the network delay requirement and improve the spectrum resource utilization efficiency.