基于改进包络图神经网络的电子元件故障预测  

Prediction of electronic component faults based on improved envelope graph neural network

作　　者：余腾龙 汪庆文 黄康 YU Tenglong;WANG Qingwen;HUANG Kang(Information and Communication Branch,State Grid Jiangxi Electric Power Company,Nanchang 330000,China)

机构地区：[1]国网江西省电力有限公司信息通信分公司,江西南昌330000

出　　处：《电子元件与材料》2024年第12期1456-1462,共7页Electronic Components And Materials

基　　金：国网江西省电力有限公司科技项目资助(52183523000F)。

摘　　要：针对电子元件故障的多样性、复杂性和隐蔽性等特点,提出一种基于改进包络图神经网络的预测方法。该方法构建电子元件运行特性序列参数集,包含静态与动态参数,输入至自注意力包络图神经网络。网络利用Time2Vec技术捕捉序列中电子元件运行时间的周期性、非周期性以及时间尺度变化等特性,从而获取序列表征结果;采用包络图卷积网络分析表征结果,深度提取电子元件运行空间上的特征信息;采用自注意力机制对该特征进行加权、拼接处理后,将结果输入全连接层中,通过sigmoid激活函数输出故障概率。测试表明,该方法能有效获取运行时序特性参数,预测故障概率,区分已发故障与潜在故障,且预测结果的锐度值均低于-0.65。Aiming at the diversity,complexity and concealment of the electronic component faults,a prediction method based on improved envelope graph neural network was proposed.A sequence parameter set of electronic components operating characteristics was constructed,including static and dynamic parameters,which were input into the self-attention envelope graph neural network.Time2Vec technology was utilized to capture the periodic,aperiodic,and time-scale changes of the running time of the electronic components within the sequence,thereby the sequence characterization results was obtained.The envelope graph convolutional network was employed to analyze and characterize these results,and the characteristic information of the electronic components running space was extracted in depth.After the features were weighted and spliced by the self-attention mechanism,the result was input into the fully connected layer,and the fault probability was output by the sigmoid activation function.The test results show that this method can effectively obtain the characteristic parameters of running time,predict the probability of failure,distinguish the existing and potential faults,and the sharpness value of the predicted results is lower than-0.65.

关 键 词：自注意力 包络图神经网络 电子元件 故障预测 运行特性序列 故障概率 

分 类 号：TP389[自动化与计算机技术—计算机系统结构]