Research On Fault Diagnosis Of Communication System Based On Bayesian Network

In the information age,the importance of the communication system is self-evident.When it fails,it needs to be checked and repaired in the shortest time,which places very high professional requirements on operation and maintenance personnel.In order to help them locate faults more quickly,an automatic fault diagnosis system was proposed,and its high efficiency and real-time performance have attracted the attention of various communication equipment companies.However,some algorithm models,such as decision trees,support vector machines,etc.,cannot do anything about the uncertain relationship between fault types and fault characteristics,and cannot achieve outstanding results.It is in such a situation that Bayesian network attracts the attention of researchers.Aiming at the problem of difficult standardization of communication system data and complex model construction based on Bayesian network,this thesis mainly conducts two parts of research: First,according to the requirements of Bayesian network for data format,the fault data of communication system is carried out.Data preprocessing,including a series of processing procedures such as mining,extraction,screening,and cleaning,has initially formed the standard sample required to construct the initial model;and based on this,the continuous feature discretization algorithm is discussed,and the K-means algorithm and Chi Merge algorithm has been improved to improve their discretization effect.Second,different from the previous Bayesian network that relies solely on basic data to build,this thesis uses the packaging method to determine the fault feature domain and conducts the Bayesian network structure learning based on the integration of expert knowledge,which greatly improves the model efficiency of the model construction.In this thesis,the effectiveness of two improved data discretization algorithms is verified by experiments.The Chi Merge algorithm improves the accuracy of naive Bayes classifier by9.32% compared with the box method.The K2 algorithm improves the efficiency of Bayesian network structure learning by 14.48% compared with the original algorithm.And the number of iterations of wrapper algorithm is reduced from more than 50 to 24 by the initial feature domain determined by expert knowledge.Finally,the diagnosis success rate of the specific model obtained by fusing the expert knowledge reached 88.67%.For the traditional fault diagnosis algorithm model,it can also meet the diagnosis requirements under the condition of missing characteristic data.The experimental results also confirmed that the Bayesian networkbased fault diagnosis model can be applied in practice to meet the needs of operation and maintenance personnel for the automatic and rapid location of faults in OLT-like communication systems.