Research On Fault Prediction Technology Of Optical Networks Based On Deep Learning

In recent years,with the continuous expansion of optical network scale and the deepening of network heterogeneity,the difficulty and cost(OpEx)of network management and maintenance have been increased.The research of fault prediction is of great significance for predicting network anomalies,improving network operation and maintenance efficiency and reducing operation and maintenance costs.However,the traditional fault prediction technology lacks comprehensive analysis means for a large number of log data and intelligent fault prediction mechanism.Deep learning technology has strong feature learning ability,and can complete accurate analysis and prediction of complex problems.It provides a new idea for improving the effectiveness of fault prediction in optical networks.Based on the monitoring data of an operator's network system,combined with DL,this paper studies the problem of accurate prediction of optical network faults under the analysis of log data characteristics.The maj or contribution of this paper includes:Firstly,participating in the design of an intelligent network management architecture,called Self-Optimizing Optical Network(SOON).The architecture is based on software-defined optical network(SDON)and integrates DL technology.The architecture realizes the interaction between controller and open source deep learning platform through private interface protocol,and provides off-line training control and online decision-making mechanism for a variety of network application scenarios.Secondly,for the data quality problems such as data loss,data discontinuity and data redundancy in the network system monitoring log,data preprocessing methods and processes are designed,which can extract effective data sets from the complicated data.Aiming at the problem of unbalanced feature distribution of original dataset,a data enhancement algorithm for gradual data and catastrophe data is designed to enhance the balance of data feature distribution and improve the efficiency of deep learning algorithm learning.Thirdly,based on the above data augmentation algorithm,a deep fully connected neural network model is designed to verify the validity of the data enhancement algorithm for balancing data feature distribution in fault prediction.The experimental results show that proper data augmentation can effectively improve the prediction effect of deep learning algorithm for alarm and non-alarm classification,and the prediction accuracy can be improved by 5.5%on the enhanced data set.At the same time,it can improve the multi-classification prediction effect of multi-fault scenarios,and improve the prediction accuracy of 4.69%on enhanced data sets.