| In recent years,with the rapid growth of global Internet traffic,the deployment of optical transport networks has continued to expand and the degree of heterogeneity has continued to deepen,posing huge challenges to the operation and maintenance of optical transport networks.Fault analysis is an important part of fault management.How to quickly and accurately analyze faults is of great significance to improving the O&M efficiency of optical transmission networks and reducing O&M costs.The traditional fault analysis technology lacks effective methods to deal with a large number of alarm data generated by faults and cannot meet the needs of intelligent optical transport network.The rapid development of deep learning algorithms in recent years provides a new solution to this problem.Based on the real data of an operator’s existing network,this paper uses big data processing methods such as data preprocessing and data mining to analyze the data.On the basis of the data analysis,a graph convolutional neural network model combined with multi-task learning is designed to Solve the problem of fault location,A recurrent neural network with embedded graph convolution is designed to solve the fault localization problem for the scenario of multiple concurrent faults.The details are as follows:First,the study of data processing methods for fault location.In view of the problems of the data collected from the existing network,such as missing data and redundant data,this paper propose strategies for data pre-processing such as data deduplication and data imputation.For the problem of that alarm correlation mining cannot locate the final fault type and fault point,this paper completes the annotation of the fault event data by combining the equipment documentation and the expert index database,and obtains a preliminary fault data set containing 36 alarm transactions.For the problems of uneven distribution of different types of fault data features in the data set and small data set,this paper proposes two data enhancement strategies based on optical transmission network topology and based on data structure.Second,Fault location technique based on deep graph convolutional neural network.For the problems of excessive fault data,non-euclidean fault data and difficult fault location in the existing network of optical transmission network,this paper designs a graph convolutional neural network model for fault location by combining multi-task learningto deal with non-Euclidean data of topological graph type.Simulation applied on self-optimizing optical network platform show that the fault localization accuracy of the graph convolutional neural network model can reach 96%,and the localization delay is between 2.9ms and 4ms.Third,the recurrent neural network fault location model with embedded graph convolution.For the problem of concurrency this paper designs a recurrent neural network model with long and short-term memory embedded in graph convolutional neural network to perform fault location.The innovative feature is that it combines the feature learning ability of the graph convolutional neural network for spatial structure information and the feature learning ability of the long-and short-term memory recurrent neural network for time series information.The simulation results show that the average subset accuracy of the topological map for fault localization when the model is optimal reaches 93.3%,the accuracy reaches 97.4%. |
