| With high bandwidth and security,atmospheric laser communication has been considered to be a key technology to solve the bottleneck problem of wireless access,and an important solution for the sixth generation(6th Generation,6G)communication.However,atmospheric laser communication is more susceptible to the influence of the communication channel than the traditional wireless radio frequency(Radio Frequency,RF)communication technology,among which atmospheric turbulence is one of the most important factors,which will cause random fluctuations in the beam power and destroy the wavefront of the transmission beam.Adaptive Optics(AO)is proved to be an effective method to restore the wavefront damaged by atmospheric turbulence and improve the received power of the signal,thus alleviating the impact of atmospheric turbulence on atmospheric laser communication.According to whether the wavefront sensor(WFS)is used,the adaptive optics technology can be divided into two types:wavefront sensor-based adaptive optics and wavefront sensor-less adaptive optics.The wavefront sensor-based adaptive optics system is complex and expensive,while the traditional wavefront sensor-less adaptive optics has poor real-time performance and unstable recovery accuracy due to the use of iteration-based control algorithms.These problems affect the application of adaptive optics in atmospheric laser communication systems directly.With the development of deep learning(DL)technology in recent years,the DL-based wavefront sensor-less adaptive optics has been widely discussed,which can avoid the iterative process of the wavefront sensor-less adaptive optics control algorithm,therefore greatly improving the real-time performance.So the DL-based wavefront sensor-less adaptive optics is expected to become the mainstream solution to the turbulence problem of atmospheric laser communication,and further promote the development of atmospheric laser communication.However,the DL-based wavefront sensor-less adaptive optics still has a certain distance from the large-scale deployment in the atmospheric laser communication system.There are three main problems:first is the difficulty of data acquisition,the data for offline training of the deep learning model needs to be as consistent with the real data in the deployment environment.The quality and quantity of data directly affect the performance of turbulent compensation;Second is the accuracy of the model,the optimization of the deep learning model should take the actual needs of the adaptive optics system into consideration,to improve the model’s ability to process turbulent data;Third is the lack of system optimization.The DL-based wavefront sensor-less adaptive optics needs to improve the deployment plan on the atmospheric laser communication system,thereby reducing costs and ensuring availability.In response to the problems above,this thesis conducts a comprehensive study on the DL-based wavefront sensor-less adaptive optics from three aspects:data acquisition,model construction,and application deployment.The main research works and innovations of this thesis are summarized as follows.1.For the problem of data acquisition,this thesis first analyzes the traditional turbulence simulation methods,including Zernike polynomial method and power spectrum method,and points out the limitations of these methods.Then we proposes a Generative Adversarial Networks(GAN)based data generation method,which can fit the distribution of the collected wavefront data so that a large amount of turbulence data in line with actual communication conditions can be generated therefore mitigating the limitations of traditional turbulence simulation methods.In order to verify the validity of the method,this thesis compares the Frechet Distance(FD)and Phase Structure Function(PSF)between the turbulent data generated by the proposed method and the real turbulent data.The results show that the proposed GAN-based method can effectively learn the real turbulence distribution.Finally,simulation also shows that the proposed method can expand the wavefront data,thereby improving the training accuracy of the deep learning model from 53%to 96%,and alleviating the data acquisition problem of the DL-based wavefront sensorless adaptive optics.2.For the problem of model construction,this thesis first analyzes the basic principles of deep learning and the typical ideas of how to construct a network model.Then a Convolutional Neural Network(CNN)based wavefront retrieval algorithm is proposed.Compared with the traditional Artificial Neural Networks(ANN)based wavefront retrieval algorithm,the simulation shows that the algorithm decreases the training loss by 14%and the root mean square of wavefront by 59%.Compared with traditional wavefront sensor-less adaptive optics control algorithm,SPGD(Stochastic Parallel Gradient Descent).It proves that the DL-based wavefront sensorless adaptive optics has better real-time performance.Then this thesis proposes a self-attention-based Phase Diversity(PD)method.This method integrates the phase diversity information into the convolutional features,which further improves the accuracy of phase recovery and addresses the ambiguity problem existing in the phase recovery process.The simulation result shows that the proposed self-attention-based Phase Diversity(PD)method can further decrease the training loss by 52%.3.For the problem of the deployment of the DL-based wavefront sensor-less adaptive optics system,this thesis proposed two application schemes for the Intensity Modulation/Direct Detection(IM/DD)system and the Orbital Angular Momentum Shift Keying(OAM-SK)system respectively.For direct modulation/direct detection laser communication systems,this thesis first analyzes the feasibility of deploying a DL-based wavefront sensor-less adaptive optics system to improve the Bit Error Rate(BER).Then this thesis proposes an online deployment method based on Reinforcement Learning(RL),which can make real-time adaptive adjustments to changes in the environment after the deployment of a DLbased wavefront sensor-less adaptive optics system and enhance the robustness of the system.The simulation results preliminarily verify that the deployment method based on reinforcement learning can resolve the change of turbulence intensity effectively.For the orbital angular momentum keying modulation system,a low-cost deployment method based on a multi-head network is proposed in this thesis.This method can learn the compensation mechanism of the adaptive optics system when the network is trained offline,thus avoiding the use of an adaptive optics system in the online deployment stage and reduce the deployment cost dramatically.The simulation results prove that the method can improve the communication performance of the orbital angular momentum keying modulation system by 21%without increasing any deployment cost. |
PDF Full Text Request