Research On Optical Performance Monitoring And Impairments Compensation Of Coherent Optical Communication Systems Based On Machine Learning

The rapid increment of emerging services such as the Internet of Things,cloud computing and big data promotes the evolution of optical network towards a large-capacity dynamic optical network with refined processing,high-speed and long-distance transmission.In order to deal with the complex and diverse service requirements,optical network needs to establish high spectral efficiency and highly reliable optical connections through real-time perception of optical physical layer transmission quality and dynamic scheduling of network resources such as modulation format.However,the dynamic optical connections greatly aggravate the complexity of optical networks,which leads to that high implementation complexity is required to effectively monitor the typical physical layer parameters such as optical signal to noise ratio(OSNR).High-speed and large-capacity transmission enabling technologies such as high baud rate,high-order modulation format and wavelength division multiplexing(WDM)have notably improved the optical network capacity,however the significant fiber nonlinearity effect caused by high launch power has become one of the key factors limiting the performance of large-capacity and long-distance networks.Although the existing digital signal processing algorithms can effectively compensate fiber nonlinearity impairments,they are still difficult to be applied in practice due to the excessive computational complexity.More than that,high baud rate transmission needs to use large bandwidth and high sampling rate digital-to-analog converter(DAC),which is difficult and costly from the perspective of technical implementation.Focusing on the above problems in coherent optical communication systems,this thesis has carried out in-depth research on the technical issues that the overcomplicated physical layer parameter monitoring schemes,the mutual restriction of the effectiveness and complexity of fiber nonlinearity compensation algorithm,and the large implementation cost of high baud rate transmission system.The main research work and innovations are as follows.1.In view of the difficulty for achieving coexistence of low complexity and nonlinearity impairments tolerance in the existing modulation format identification schemes,a low-complexity and nonlinearity-tolerant modulation format identification scheme based on random forest is proposed in this thesis.Benefiting from the simple binary tree structure and the swarm intelligence achieved by decision tree integration,random forest algorithm can significantly reduce the computational complexity while effectively overcoming the adverse effects of nonlinearity impairments on the feature distribution of modulation format.16 GBaud polarization multiplexing-4/8/16/32/64 quadrature amplitude modulation(PM-4/8/16/32/64QAM)three-wavelength WDM coherent optical communication system simulation results indicate that when the optical launch power of PM-4QAM,PM-8/16QAM and PM-32/64QAM is as high as 12,7 and 6 dBm,that is,11,6 and 5 dB greater than the optimal launch power,the proposed scheme can still achieve 100%modulation identification accuracy.Moreover,only 30 decision trees with a depth of 5 are required for random forest,and its computational complexity is at least an order of magnitude lower than that of deep neural network with comparable identification accuracy.In addition,in order to further verify the effectiveness of the modulation format identification model obtained during the simulation training process,16 GBaud PM-4/16/32QAM three-wavelength WDM offline experimental systems are conducted.Experimental results show that the proposed scheme can still achieve 100%identification accuracy when the launch power is at least 4 dB greater than the optimal launch power.2.The joint monitoring of OSNR and modulation format is conducive to the achievement of cost-effective optical fiber link quality diagnosis and modulation format adaptive digital signal processing in coherent receivers,however most of the existing joint monitoring schemes require complex monitoring algorithms with high computational complexity.To cope with this issue,this thesis proposes a low-complexity and joint OSNR estimation and modulation format identification monitoring scheme through the use of random forest.The many decision trees integrated in random forest only require relatively fewer comparative calculations with lower computational complexity,which can not only operate in parallel mode,but also achieve the competitive monitoring accuracy by leveraging the group majority voting strategy and averaging tactic between the decision trees.The comprehensive simulation results of 16 GBaud PM-4/8/16/32/64QAM system show that the averaged mean absolute errors(MAE)of OSNR estimation for the proposed joint monitoring scheme are 0.24,0.29,0.31,0.38 and 0.66 dB,respectively.Furthermore,when the OSNR is as low as 5.12,7.45,10.74,15.15 and 18.22 dB,that is,it is at least 1.5 dB smaller than the soft decision forward error correction(SD-FEC)threshold,modulation format identification accuracy of the proposed joint monitoring scheme can still be maintained at 100%.Compared with algorithms such as support vector machines that require thousands of multiplication operations,the proposed random forest-based joint monitoring scheme can reduce the computational complexity by more than an order of magnitude while achieving equivalent or even better-monitoring performance.In the meanwhile,16 GBaud PM-4/16/32QAM offline experimental results show that not only the averaged MAE of OSNR estimation is smaller than that of support vector machines and deep neural networks,but also the modulation format identification accuracy can still be maintained at 100%when the OSNR is lower than the SD-FEC threshold.3.Fiber nonlinearity impairments compensation independent of specific system parameters can be effectively achieved by the combination of deep neural network and perturbation method.However,the memoryless feedforward structure makes deep neural network subject to the problem that the performance and computational complexity of nonlinearity impairments compensation are restricted each other.In response to the above issue,this thesis proposes a low-complexity nonlinearity impairments compensation scheme enabled by simple recurrent neural network(SRNN).By taking advantage of the time memory of SRNN,the proposed SRNN-based scheme can significantly reduce the number of triplets,which can reflect the characteristics of fiber nonlinearity impairments,required by the equalization process.Moreover,compared with existing solutions based on deep fully connected neural networks,the computational complexity of the proposed SRNN-based scheme is reduced by about half and its training complexity can be reduced by more than one time.The offline experimental results of three-wavelength 16 GBaud PM-16QAM coherent transmission system indicate that the proposed scheme can increase the optimal launch power by 1 dB and the Q-factor by 0.49 dB.Numerical simulation results of 30 GBaud PM-16QAM coherent optical communication system demonstrate that compared to the case of only compensating for linear impairments,the proposed SRNN-based scheme under single-wavelength and three-wavelength WDM systems can improve the optimal launch power by 1 dB,and the Q-factor can be increased by 0.59 dB and 0.38 dB respectively.Meanwhile,the maximum transmission distance can be increased by 270 km and 210 km under the threshold of hard decision forward error correction(HD-FEC).4.In order to break through the limitation of high-cost digital-to-analog converter(DAC)with high sampling rate and large bandwidth on the implementation of high baud rate/Tbps level coherent optical transmission system,this thesis designs a novel single-channel 1 Tbps coherent optical transmission system solution based on analog subcarrier multiplexing(ASCM).ASCM scheme can significantly reduce the requirements of the sampling rate and bandwidth for DAC while maintaining the advantages of subcarrier multiplexing such as nonlinearity tolerance enhancement.Combining with the K-nearest neighbor(KNN)algorithm,the fiber nonlinearity tolerance of ASCM can be further improved.The simulation results of 16×8 GBaud PM-16QAM ASCM system show that compared to the 128 GBaud PM-16QAM single-carrier system,ASCM system can increase the optimal launch power by 1 dB and Q-factor by 2.02 dB.Under the equalization of the KNN algorithm,the Q-factor can be increased by an additional 0.46 dB.Meanwhile,the maximum transmission distance of the ASCM system under the HD-FEC and SD-FEC thresholds can be increased by 800 km and 900 km,respectively.After KNN equalization,the maximum transmission distance of single-carrier and ASCM systems under the HD-FEC threshold can be increased by an additional 200 km.