Research On Optical Performance Monitoring Based On Machine Learning

With the explosive growth of global traffic demand,there is an urgent need to improve the bandwidth and speed of communication systems,and optical networks are facing significant challenges.Coherent optical communication systems have emerged as the mainstream of high-speed optical communication systems due to their higher sensitivity,higher communication capacity,and longer transmission distance.To ensure transmission quality and detect damage or interference in optical communication systems,optical performance monitoring(OPM)is necessary for coherent optical communication systems.With the development of the new generation of optical networks towards flexibility and intelligence,it is necessary to obtain transmission information at different nodes in real-time to dynamically allocate network resources and maximize efficiency.Therefore,OPM is crucial in this context.In order to learn the transmission quality of the signal,compensate for signal damage and adapt to the new generation of the intelligent optical networks,it is necessary to monitor the modulation format and OSNR of the system.Traditional OPM monitoring methods have low accuracy,require manual threshold setting,and highly invasive,making it difficult to adapt to the new generation of optical communication systems.To overcome these limitations,this paper proposes two machine learning-based OPM methods that can achieve optical performance monitoring more quickly,efficiently,and automatically.The first method is OSNR estimation based on neural networks,which is optimized on the traditional amplitude histogram-based OSNR estimation method to improve the accuracy of recognition.The second method is modulation format recognition and OSNR estimation based on small sample learning,which can achieve good recognition results without requiring a large number of samples to train the model.The neural network-based OSNR estimation method uses amplitude histograms as training objects,and we introduce smoothing processing and additional statistical information to regress the system’s OSNR.To demonstrate the effectiveness of this approach,QPSK,16 QAM,and 64 QAM coherent optical communication systems were built,and the signals were collected in digital signal processing,after that amplitude values will be calculated.Amplitude histograms were generated based on the probability distribution of the amplitude values,and the histograms were smoothed to make their distribution more regular.Additional statistical information was then added,and the smoothed amplitude histogram and statistical information were trained as input data for the neural network to ultimately obtain a regression estimate of OSNR.Experimental results show that the introduction of smoothing processing and additional statistical information can reduce the error of OSNR estimates,and the final OSNR estimates for QPSK,16 QAM,and 64 QAM modulation formats are lower than 0.14 dB,0.25 dB,and 0.97 dB respectively.Then we proposed a method for recognizing modulation formats and estimating optical signal-to-noise ratio(OSNR)based on small sample learning.Our approach uses constellation images as training objects,and employs algorithms to classify images of different modulation formats and OSNR.Specifically,we built a coherent optical communication system with two modulation modes(PSK and QAM)and ten modulation formats(BPSK,QPSK,8PSK,16 PSK,8QAM,star 16 QAM,16QAM,32 QAM,64QAM,and 128QAM).We collected constellation images generated after digital signal processing and extracted feature vectors using an embedded network.The training based on these feature vectors allowed us to achieve 100% recognition rate for all modulation formats and 100% recognition rate for OSNR estimation except128 QAM modulation format.The innovation of this work are as follows:(1)An improved method for traditional neural network based OSNR estimation is proposed.The amplitude histogram is extracted from amplitude information,then we introduced smoothing processing and additional statistical information to the system,then we regress the OSNR value of coherent optical communication systems,achieved good results.(2)A modulation format recognition and OSNR estimation method based on fewshot learning algorithm is proposed.Through a metric based meta learning algorithm,only a small number of constellation samples are used for classification,and good results are achieved.(3)The proposed modulation format recognition and OSNR estimation methods based on few-shot learning algorithm is migratable,it saves training time,improves efficiency,and is suitable for multiple scenarios,reducing costs.