Theoretical And Technical Research On Optical Signal Transmission Based On Generative Adversarial Networks

With the continuous development of new technologies such as 5G,Internet of Things and cloud computing,the optical communication system which plays a fundamental role in big data transmission,needs higher transmission performance to better adapt to the rapid development of communication transmission rate and Internet data traffic.Therefore,it is a necessary process to introduce new technology and realize ultra-high speed and large capacity by studying the related theory and technology of optical transmission network.In recent years,convolutional neural networks(CNN)in deep learning(DL)has been widely used in optical signal performance monitoring and physical impairments diagnosis,while generative adversarial networks(GAN)known as the mainstream technology in the field of computer vision,plays an important role in the direction of image generation and data augmentation.Therefore,combining GAN with the signal processing and data enhancement in optical transmission system can effectively solve the problems of lack of training data and weak model generalization in traditional optical communication.Based on the strong learning ability of DL and the powerful application of GAN in the field of image processing,this paper focuses on the data augmentation of eye-diagrams and constellations diagrams in optical transmission system.The main work and innovations of this paper are as follows:Firstly,because the existing application technology of DL combined with optical signal mainly focuses on the neural network model structure,which relies on large amount of data and strong computing power,a scheme about optical signal image generation based on GAN is proposed.It verifies the feasibility of applying GAN to optical signal for image data augmentation.Secondly,aiming at the problem that CNN lacks large-scale and high-quality datasets during model training,an image generation scheme of eye-diagrams based on deep convolutional generative adversarial network(DCGAN)is proposed.The simulation results show that,this scheme can generate new images similar to the real samples based on the eye-diagrams dataset of PAM4 in the intensity modulation direct detection systems,and the distinguish correct rate of the discriminator can be close to 0.5 after several rounds of iterative training,which shows that the ideal eye-diagrams generation effect can be obtained.It verifies the feasibility and efficiency of the model using data enhancement method.Thirdly,to solve the shortcomings of the above-mentioned image generation scheme based on DCGAN,which is too free and uncontrollable to generate images,an image augmentation scheme based on conditional generative adversarial network(CGAN)is proposed.In the data enhancement of the eye-diagrams of PAM4 and the constellation diagrams of PM-16QAM,different impairment types and different transmission distances are taken as model conditions to realize the data augmentation.The simulation results show that,by comparing CNN classification models using original datasets and expanded datasets respectively,it is found that the classification ability of the augmented datasets improves about 5%than that of datasest before augementation,which greatly improves the generalization of classification model and effectively solves the problem of insufficient training sample dataset of algorithm.