概率分布表插值的离散贝叶斯网络故障诊断算法

Discrete Bayesian Network Fault Diagnosis Algorithm Based on Interpolation of Probability Distribution Table

针对离散贝叶斯网络(BN)进行故障诊断时概率统计因子不充分的问题,提出了基于JS散度约束的二维插值算法对网络条件概率分布表进行插值。首先,根据故障数据分析故障特征之间的依赖关系,以条件互信息作为节点的连接权重确立贝叶斯网络拓扑结构;其次,将故障特征进行离散化处理,用标记的故障数据学习初始的贝叶斯网络概率分布表;最后,选择插值函数的类型和参数对所得的概率分布表进行二维插值,解决由于故障数据不完备而导致的概率分布表中概率统计因子不连续以及概率值零值较多的问题。通过实验数据仿真对比该方法比没有插值的网络诊断准确率提高11.68%。

Aiming at the problem of insufficient probability and statistical factors in the fault diagnosis of discrete Bayesian network(BN),a two-dimensional interpolation algorithm based on JS divergence constraints is proposed to interpolate the network conditional probability distribution table.First,analyze the dependency relationship between the fault features according to the fault data,and use the conditional mutual information as the connection weight of the nodes to establish the Bayesian network topology;then,the fault features are discretized,and the initial Bayesian is learned from the labeled fault data.Network probability distribution table;finally select the type and parameters of the interpolation function to perform two-dimensional interpolation on the resulting probability distribution table to solve the discontinuity of the probability statistical factors in the probability distribution table caused by incomplete fault data and the probability value of zero problem.Compared with the network diagnosis accuracy rate without interpolation,the accuracy of this method is improved by 11.68%by comparing the experimental data simulation.