基于高斯过程分类器的变压器故障诊断

Fault Diagnosis of Transformers Based on Gaussian Process Classifier

构建了基于拉普拉斯近似方法的高斯过程分类器(LGPC)。LGPC可自行优化超参数,以概率形式输出分类结果,便于问题的不确定性分析,从而克服SVM规则化系数、核函数参数确定困难等局限。在用典型分类数据验证LGPC在分类性能方面优于SVM的基础上,提出了基于LGPC的变压器故障诊断方法,并给出了其具体实现方法。通过工程实例验证了均值函数采用常函数、协方差函数采用全平方指数函数、似然函数采用误差函数时,故障诊断的正确率较高。同基于SVM的故障诊断方法相比,本文所提方法可以取得更高的故障诊断正确率,具有可行性和推广性。

Gaussian process classifier based on Laplace approximation method(LGPC) is constructed.It can optimize the hyper parameters of the LGPC automatically,output classification results in probability,and be convenient to analyze problems’ uncertainty.Therefore,LGPC can overcome the inherent limitations of SVM whose regularization factors and kernel function parameters are difficult to determine.In this paper,performance of LGPC is analyzed and validated by typical classification datasets,and transformer fault diagnosing method based on LGPC is presented and described in details.Experimental results show that the diagnosing correctness ratios are higher when mean function adopts a constant function,covariance function adapts a full square exponential function and likelihood function adopts an error function.Compared with methods based on SVM,the proposed method has higher classification accuracy,which proves it is effective.