Optical Network Performance Prediction Based On Deep Learning

Recently,some new optical communication technologies such as high-order modulation,wavelength division multiplexing are widely used in optical fiber communication systems.However,more and more problems appeared such as complex optical network structure,increased signal damage,which seriously affects the quality of optical network transmission.Therefore,optical performance monitoring technology,as one of the key technologies of optical network transmission,needs to be improved and updated continuously.This thesis focuses on optical performance monitoring technology,studies the constellation diagram recovered from the coherent optical module of the optical transmission system terminal.Establishing an intelligent prediction model of terminal optical performance based on deep learning,and realizes accurate calculation of performance parameters such as optical signal-to-noise ratio and chromatic dispersion.It is of great significance for ensuring the accuracy of signal transmission,the stability of optical network transmission,and optimizing network distribution.The main research work and innovations of this thesis are as follows:(1)This thesis proposes a computing approach based on attention mechanism for optical signal-to-noise ratio to study the changes in the performance of optical fiber links such as device aging,temperature changes,and man-made continuity operations.Through data collection and experimental demonstration,the calculation error of the optical signal-to-noise ratio monitoring and prediction algorithm is less than 0.5dB,and does not increase with the change of transmission distance and fiber power.It shows that the optical signal-to-noise ratio monitoring and prediction algorithm based on the attention mechanism has high calculation accuracy,and it is suitable for a variety of transmission distances and fiber input powers.(2)This thesis proposes a prediction model of optical performance parameters based on multi-task learning,to predict some optical performance parameters such as optical signal-tonoise ratio,bit error rate,and dispersion are predicted at the same time.The optical signal-tonoise ratio prediction error is reduced to 0.1dB,the bit error rate prediction error is 0.7,the dispersion prediction error is 0.1 ps/(nm?km),and the calculation error does not increase with the transmission distance.It shows that the multi-task-based optical performance parameter prediction algorithm improves the prediction accuracy compared with the single-task prediction algorithm,and it is suitable for a variety of transmission distances,thus can predict the changes in optical link performance comprehensively.(3)This thesis proposes a new optical module performance parameter prediction model and training program based on transfer learning.Transferring the optical signal-to-noise ratio prediction model based on the attention mechanism in the first research work to another new optical module prediction model.Labeling one-fifth of the sample data on the new optical module,the model training time is about one-tenth the time of non-migration training,and the optical signal-to-noise ratio calculation error is 0.349dB.It shows that the new optical module performance parameter prediction model and training scheme based on migration learning have high calculation accuracy,less time and cost saving for model training.(4)This thesis proposes an intelligent model distribution method and distribution system.Based on prospect theory to quantify the gains and losses of the allocation plan,reasonably allocate the intelligent model to the cloud platform and device side,give full play to the cloud platform's computing power,large capacity,and the device side's low latency,and high security.It helps to achieve a high degree of synergy between the optical network and the cloud.