Research Of Optical Performance Monitoring Based On Deep Learning

With the prosperous developing of Artificial Intelligence(AI),it has taken on very vital tasks in all walks of life.Machine Learning(ML),as a representative of the new generation of the powerful intelligent analysis tools,has become the one of indispensable research topics.Notably,due to the massive popularity of Internet,Big Data has been one of the vital labels for this era.But most of the data are redundant and useless.In this case,what we need is to effectively extract the most critical and core data features from that for analysis and prediction.If we utilize traditional machine learning methods,such as Knearest Neighbor or Support Vector Machine,etc.,it will be an extremely difficult thing to tackle such a huge amount of data analytical processing task.Therefore,Deep Learning(DL)is on the stage of history.DL is an exceptional technique for implementing ML.It enables ML to achieve numerous applications,which is mainly reflected in the advancing innovations of neural network itself.And in the field of optical fiber communication,the monitoring on system parameters is a really pressing concern now,which kind of task is also called Optical Performance Monitoring(OPM).But so far,there are relatively few studies on the use of DL in this area to try to solve this problem.However,for the case of the limitation on the existing OSNR traditional techniques' monitoring accuracy and efficiency,and the required equipment complexity and labor costs are significantly high,the traditional monitoring schemes based on communications itself has gradually been replaced by DL,due to its powerful feature fitting,parameter prediction,excellent monitoring effect and training speed,etc.,so it has also received extensive attention and research from both the academic and industry circles.This thesis is mainly based on both the simulation and experiment schemes,which focus on parameters-prediction on the two types of mainstream optical fiber communication systems.We conducted the researches on high-accuracy monitoring and also the use of transfer learning methods to aid the DNN for the performance improvement.Firstly,the combination of deep learning and optical performance monitoring was used to improve the speed and accuracy of parameter monitoring.Meanwhile,the monitoring tolerance of DNN was also tested when the signal was under different external interference circumstance.Then we proposed a method using transfer learning to assist deep learning to solve this problem and to speed up the remodeling of DNN in multiple application scenarios.The experimental results also show that our method can successfully reduce the required amount of training data and speed up its training speed.Finally,the autoencoder is also used to replace the traditional amplitude histogram method for data feature extraction.The test results found that the scheme can indeed complete the required performance monitoring tasks to a certain extent.