Theoretical And Technical Research On Intelligent Eye Diagram Analysis Based On Deep Learning In Optical Communication

Medium to short optical transmission technologies are widely used in optical access and optical interconnection systems,providing a promising solution for the emerging Internet services with increasing bandwidth and speed requirements.As the future of the network moves toward dynamic and complex format features,it is critical to perform a full range of performance monitoring of the system to ensure stable network operation.In the Intensity modulation direct detection(IM-DD)system for medium and short-distance transmission systems,eye diagram monitoring shows a greater advantage.The machine learning-based monitoring scheme currently under study,with its innate advantages in dealing with nonlinear problems,achieves better eye diagram monitoring than traditional oscilloscopes.However,since traditional machine learning(ML)algorithms require artificial feature extraction and may cause loss of information during feature construction,they are still not suitable for the development of flexible dynamic networks in the future.Deep learning(DL)technology provides a high-precision and intelligent approach to eye performance monitoring with its automatic feature extraction capability and excellent performance.This paper proposes several technical solutions based on deep learning technology for intelligent monitoring of eye diagrams.The main innovations of the paper are as follows:Firstly,aiming at the problem that the traditional machine learning eye diagram monitoring method lacks automatic extraction ability and poor performance,a deep learning intelligent eye monitoring scheme based on convolutional neural network(CNN)is proposed.The experimental results show that the scheme can realize OSNR estimation and modulation format identification under various modulation formats of OOK,NRZ-OOK,DPSK and PAM4,and fiber link analysis and Q factor estimation in OOK and PAM4 modulation formats.Compared with a variety of Traditional ML algorithms,100%estimation accuracy can be achieved at 1dB interval,which is significantly better than other traditional ML algorithms.Secondly,aiming at the problem that the above-mentioned CNN-based eye diagram monitoring scheme takes a long time to train and the model is difficult to reuse,an intelligent eye diagram monitoring scheme based on deep transfer learning is proposed,and two implementations of Fine-tune and Frozen are established respectively.The experimental results show that the scheme can effectively reuse the eye image feature information of the e xisting model,and both implementations significantly shorten the training time of other related eye monitoring tasks,and the performance of Frozen is more prominent.Compared to the no-transfer models,the Frozen-based transfer method reduced training time by at least 96.84%and 97.69%in the OOK and PAM4 formats,respectively.Thirdly,focusing on the problem of repeated training and low monitoring efficiency in the single task eye diagram monitoring scheme based on transfer learning,a multi-task intelligent eye monitoring scheme based on deep transfer learning is proposed.The experimental results show that the multi-task learning mechanism can realize the joint monitoring of multiple eye parameters in parallel,avoiding the repeated training of each model and significantly improving the efficiency of eye diagram monitoring.Compared with the single task transfer model,the Frozen-based multi-tasking transfer method realizes parallel monitoring of multiple tasks with only one task training duration in the OOK and PAM4 formats.