Research On Performance Monitoring Of Optical Fiber Communication System Based On Deep Learning

With the rise of the cloud computing,the internet of things(IoT),the fifthgeneration of mobile communication(5G)and artificial intelligence(AI),the massive data transmission demand is promoting the development of optical fiber communication network towards the direction of high-capacity,dynamization,heterogeneity,flexibility and intellectualization.Meanwhile,the AI technology which mainly represented by deep learning(DL)is becoming the beginning of the fourth industrial revolution,and gradually promotes the human society into the era of intelligence.The combination of AI and optical fiber communication network is the.only way to promote the intelligent development of the optical communication network.Using deep learning to monitor the parameters which reflecting the operation status of the optical network is an important guarantee for the stable operation and good management of optical network.However,the monitoring technologies based on deep learning for optical network are still immature.It has the problems of poor monitoring accuracy,low monitoring reliability and poor adaptability,which seriously restrict the intelligent development of optical network.This dissertation studies the several key problems such as the multi-task monitoring,the improvement of monitoring accuracy,reliability and adaptability.Based on deep learning and the analysis of monitoring tasks’ characteristics,our work lays a good foundation for the realization of the accurate,reliable and adaptable monitoring.The main research work and innovation achievements of this dissertation are summarized as follows:(1)In order to solve the problem that the monitoring accuracy of optical performance monitoring(OPM)and bit-rate/modulation format identification(BRMFI)is poor,a joint OPM and BR-MFI algorithm based on multi-task learning is proposed.By designing the multi-task learning model and using the model’s ability of automatic feature extraction and feature sharing,the two tasks can promote each other’s performance.The simulation results show that compared with the principal component analysis(PCA)based OPM and BR-MFI algorithm,the monitoring accuracy has been significantly improved.(2)In order to solve the problem that the monitoring of OPM and BR-MFI is unreliable,a parallel monitoring framework is proposed firstly.Then,a skipconnected generative adversarial network(GAN)algorithm is proposed in the judgement module of the parallel monitoring framework.In the paralle monitoring framework,the newly added judgement module works in paralle with the data analysis module,which improves the monitoring efficiency.In the judgement module,the skip connected GAN algorithm is designed to filter the illegal data by learning the distribution of the legal data.The simulation results show that the algorithm can identify and filter the illegal data,which improves the reliability of the monitoring.(3)In order to solve the problem that current OPM and BR-MFI algorithm lack the ability to adapt the dynamic optical network environment,a transfer-learning based joint OPM and BR-MFI algorithm is proposed.The transfer-learning method can transfer the learned knowledge from the labeled data under an invariant optical network environment to the unlabeled data under a dynamic optical network environment.During the training phase,the maximum mean discrepancy(MMD)as one of the training constraints,is applied to match the features extracted from the source and target domains.The simulation results show that the algorithm has the adaptability of the dynamic environment.(4)Considering the mode division multiplexing(MDM)scenario,and in order to solve the accuracy deterioration of modal decomposition(MD)which caused by modal ambiguity,a convolution neural network(CNN)combined with the near and far-field images reconstruction losses is proposed.The model is trained with the combination of a modal coefficient and two reconstruction losses(near-field and far-filed intensity reconstruction losses),so as to mitigate the problem of modal ambiguity.The simulation results show that the algorithm can attain accurate modal coefficient for all mode cases.The experimental results show that the accuracy of MD on the measured near-field image is reduced,which may be caused by image size adjustment and mode crosstalk.The above research results provide feasible ideas and solutions for improving the accuracy,reliability and adaptability of monitoring,and provide important technical support for the intelligent development of optical networks.