样本规范化在故障诊断神经网络方法中的应用  

Sample Normalization in Fault Diagnosis Based on Neural Network Method

作　　者：禹建丽[1] 徐云龙[1] 张黎[1] 陈洪根[1] 

机构地区：[1]郑州航空工业管理学院管理工程学院,郑州450046

出　　处：《河南科学》2016年第12期1961-1967,共7页Henan Science

基　　金：国家自然科学基金项目(U1404702);航空科学基金项目(2014ZG55021);河南省自然科学基金项目(162102210083);河南省高等学校重点科研项目(17A630069);郑州航院大学生科技创新项目(Y20150104)

摘　　要：在神经网络模型中,训练样本的不同特征变量(输入变量)反映响应变量(输出变量)的灵敏程度有一定的差别,若将各特征变量直接作为神经网络的输入,则会湮没一些弱小量可能含有的有效信息.采用油中溶解气体分析法建立以变压器油中溶解气体含量为输入,变压器故障类型为输出的BP神经网络变压器故障诊断模型,分别运用最大值最小值规范化、一般浓度规范化、特征浓度规范化三种方法对训练样本进行规范化.研究结果表明,不同样本规范化方法对故障诊断效果的影响显著,运用特征浓度规范化进行规范化处理要优于其他方法.In the model of neural network, there are a certain difference in sensitivity that the different characteristics variables of training samples reflect the responses of variables. If we put the characteristic variables as the input to the neural network directly, some small amount that may contain useful information will be lost. This paper establishes a BP neural network model for transformer the fault diagnosis based on the dissolved gas analysis, takes the gas content in transformer oil as the input and the transformer fault types as the output. Using the three methods： the min-max normalization, the general concentration normalization and the characteristics concentration normalization, the training samples are standardized respectively. The research results show that the different sample normalization method has significant effects on the performance of the fault diagnosis model, and it is better than other methods to standardize processing by using characteristic concentration normalization.

关 键 词：变压器 BP神经网络 规范化 故障诊断 

分 类 号：TP277[自动化与计算机技术—检测技术与自动化装置]