线性分类器与BP网络联合诊断变压器故障

Diagnosis of Failed Transformer Using the Combination of Linear Classifier and BP Network

油中溶解气体分析(DGA)是目前电力充油设备潜伏性故障诊断的重要手段。为了克服传统BP网络及其改进诊断算法所具有的隐层节点数多、收敛时间长的缺陷,减少算法运算量及提高变压器故障诊断的正确率,提出了一种新的诊断算法:线性分类器-BP神经网络(LC-BP)故障辨识方法。通过对变压器大量过热和放电两类典型故障数据的研究,发现其DGA故障数据的特征空间线性可分且分离度较好。基于以上特性,先用线性分类器诊断过热和放电故障,然后利用两个小型BP网络分别进行进一步诊断,得到最终诊断结果。实验结果表明,提出的LC-BP算法具有良好的分类能力,故障诊断的正确率达到94%,且网络结构简单,运算量小,从而为变压器的故障诊断提供了一条新的有效途径。

The dissolved gases analysis （ DGA ） problem of to diagnosing the internal faults of the electrical devices filled with oil is discussed. To overcome the disadvantages of the conventional BP neural network, and to increase the diagnostic correctness rate while lessen the calculation, an algorithm based on linear classifier and BP neural network （LC-BP） is presented. Based on the analysis of DGA data from the failed transformers, the characteristic space of the DGA data can be divided into too-hot space and discharge space efficiently. The faults of whether too-hot or discharge are diagnosed by the linear classifier firstly. Then two little BP neural networks are utilized to determine the eventual faults. The simulation results show that the proposed LC-BP algorithm proposed has good ability of classification, simple structure and little calculation, while the diagnostic correctness rate reaches at 94 %.