Research On Optical Signal-to-noise Ratio Estimation And Modulation Format Identification Technologies Based On Deep Learning

With the rapid development of technologies such as big data and artificial intelligence,data traffic has seen a surge in growth.Fiber optic communication is widely used due to its enormous transmission bandwidth,large communication capacity,low transmission loss,anti electromagnetic interference,and good confidentiality.In 2021,China’s annual mobile internet user access traffic exceeded 200 billion GB.At present,over 97% of China’s information is transmitted through fiber optic cables,with a total length of over 550 million kilometers.In order to improve spectral efficiency,elastic optical networks have been proposed.It can flexibly adjust modulation format,transmission rate,etc.at the transmitting end based on user needs and real-time link status.However,at the receiving end,digital signal processing(DSP)needs to use the modulation format as prior information and select a compensation algorithm that matches the modulation format to achieve precise compensation.In addition,optical signal to noise ratio(OSNR)directly reflects the degree of signal interference from channel noise and is directly related to bit error rate.It is the key to fault diagnosis and localization in optical transmission networks.Therefore,how to accurately identify modulation formats and estimate OSNR without prior information has become an urgent problem in optical communication system.The existing modulation format recognition and OSNR estimation schemes based on deep learning have problems such as low intelligence,slow task convergence,and insufficient generalization ability.Focusing on these problems,this article conducts research on modulation format identification(MFI)and OSNR estimation algorithms based on deep learning.The specific work content is as follows:1.A novel algorithm based on adaptive multi-task learning(AMTL)combining MFI and OSNR estimation is proposed to address the low level of intelligence in previous multi task monitoring schemes.This algorithm performs feature extraction on the constellation extracted by the receiver,accurately distinguishing dual polarization-quadrature phase shift keying(DP-QPSK)and multi-level quadrature amplitude modulation(M-QAM),and estimating OSNR within the commonly used range.In order to verify the feasibility of this scheme,an optical transmission simulation system with a transmission rate of32 G Baud was built.The OSNR estimation task was defined as classification and regression problems respectively by two adaptive multi-task learning models.The results show that classified convolutional neural network(CNN)can accurately recognize MFs and OSNRs 100%;Combined CNN can not only identify MFs 100%,but also estimate OSNR with an average absolute error(MAE)of only 0.50 d B and a root mean square error(RMSE)of 0.66 d B.Compared with adaptive multi-task multi-layer perceptrons(AMT-MLP)and adaptive multi-task recurrent neural network(AMT-RNN),adaptive multi-task convolutional neural network(AMT-CNN)has faster convergence speed and smaller model complexity.In addition,the impact of fiber nonlinearity on the adaptive multitask learning model was studied,and the proposed algorithm was verified to have a certain tolerance for nonlinearity.2.An algorithm based on adaptive feature fusion is proposed to address the problem of slow convergence caused by negative transfer between related tasks.This algorithm is used to accelerate the parallel optimization speed of multiple tasks on small sample sets.To verify the effectiveness of the algorithm,the OSNR estimation was treated as a classification problem and a regression problem,and a CNN with the same backbone was constructed Three ablation models,including feature fusion-based convolutional neural network(FF-CNN)and adaptive feature fusion-based convolutional neural network(AFF-CNN),are trained and tested on the same dataset.The results show that the algorithm achieves the highest accuracy and the smallest error,with the recognition accuracy of modulation format and OSNR for different types of AFF-CNN reaching 100% and 98.6%,respectively;The combined AFF-CNN can also recognize modulation format types 100%,with a MAE of only 0.33 d B and RMSE of 0.46 d B for OSNR estimation.This AFF algorithm has the effect of accelerating the convergence speed of MFI and OSNR estimation tasks,and at the same time,its prediction performance is more stable on small sample sets.This algorithm hardly increases the complexity of the network,and to some extent,it is more in line with the performance requirements of actual systems.In addition,the influence of fiber nonlinearity on three ablation models was studied from the perspectives of transmission power and transmission distance,and the proposed algorithm was verified to have a certain tolerance for nonlinearity.3.To solve the problem of insufficient generalization ability of optical signal-to-noise ratio estimation models on small sample sets,a multi-task model-agnostic-meta-learning algorithm is proposed for joint MFI and OSNR estimation.This algorithm can complete learning tasks with very few samples when transferring models.In order to verify the performance of the algorithm,three ablation algorithms were designed,including AMTL,single-task model-agnostic-meta-learning(ST-MAML),and multi-task model-agnostic-meta-earning(MT-MAML).Firstly,the influence of hyperparameter on the performance of the algorithm is analyzed by changing the proportion of support set and the number of gradient descent steps.The results show that MT-MAML can achieve the optimal results by only using a single step gradient descent,and the algorithm complexity is much lower than that of the comparison algorithm;When there is only one fine-tuning sample per category in the support set,the MFI accuracy of MT-MAML is 100%,and the OSNR estimation accuracy can reach 96.4%,which is about 36% higher than the ST-MAML algorithm;When there are 5 samples in each category of the support set,the OSNR estimation accuracy reaches98.6%,which is about 10% higher than the ST-MAML algorithm.Finally,a generalization study was conducted on the proposed algorithm to demonstrate that the proposed model can quickly adapt to the data distribution in new scenarios,achieving an accuracy of over 90% within two epochs.The proposed multi task meta learning scheme can quickly adapt to new tasks and data,and has excellent generalization ability.It is highly attractive for optical performance monitoring devices in isomerized and dynamic optical networks.