Study On Root Alarm Analysis Technology Based On Alarm Vectorization In Optical Network

As move into a new era,the country has emphasized the requirements for digital economy development planning at the beginning of the 14th Five-Year Plan.In 2022,it also drew a grand blueprint for building a new computing power network system,and decided to launch the great strategic project of "National Computing Network to Synergize East and West".High-speed network construction is the basis for achieving balanced layout of computing power.Optical network,with its characteristics of low delay and large capacity,is very suitable for the network required by "National Computing Network to Synergize East and West".Therefore,further expanding the scale of optical network will be an important proposition of the times.As an important infrastructure in the new era,optical network cannot be ignored due to network failures caused by time,environment and even human factors while its scale is expanding,which requires active and effective fault management measures.As an effective means of fault management,the core of alarm analysis is to find the root alarm that can locate the fault.It is difficult to identify and analyze the root alarm from a large number of complex alarm information.Faced with the problems of limited accuracy and low efficiency in the analysis based on artificial recognition,the intelligent technology represented by neural network provides more help for the fault management of optical network.This thesis focuses on the analysis and identification of root alarm in optical network fault management in a neural network perspective.The main minds are as follows.First,aiming at the problem that the neural network cannot handle non-numerical alarm symbols,this paper decides to use low-dimensional,dense distributed vectors to represent alarms.On this basis,in order to obtain more scientific and effective alarm representation,this thesis proposes a knowledge-enhanced alarm vectorization scheme.It uses expert knowledge text in the alarm field,extracts knowledge features through natural language processing technology,and generates alarm representation to realize knowledge enhancement.The experimental results show that the alarm representation generated by knowledge enhancement alarm vectorization can not only be easily integrated into the downstream model,but also achieve the similarity measurement between alarms.More importantly,taking alarm knowledge as algorithm priori can improve the effect of the downstream model to identify the root alarm.Secondly,in order to find the root alarm from the alarm transaction,this dissertation constructs a root alarm identification model that is compatible with knowledge-enhanced alarm vectorization.It consists of three modules,which can encode the sequence characteristics of the alarm transaction.Its core is a matching mechanism,that is,to determine the root alarm by calculating the dot product similarity between the alarm vectors and the transaction representation.The proposed model not only solves the target optimization problem of identifying the root cause alarm based on neural network,but also the experimental results show that the model can complete the identification task with high F1-score(95%)and hit rate(HR@3=1.0),in addition,verify the effectiveness of alarm knowledge priori for identifying the root cause,and can also quantify the correlation degrees between alarms.