基于深度森林的变压器故障诊断方法

Transformer Fault Diagnosis Method Based on Deep Forest

针对变压器传统故障诊断方法准确率低以及难以有效处理故障特征信息的问题,提出了一种基于深度森林(deep forest,DF)的变压器故障诊断方法。考虑到油中溶解的特征气体与故障类型的相关性,首先将油中溶解特征气体的无编码比值作为深度森林模型的特征参量,再通过多粒度扫描提取更多的变压器故障特征信息,最后经过级联森林的训练达到诊断效果最优。通过算例分析了不同特征参量和不同样本规模下DF模型的诊断效果,并与多类传统诊断方法相比较。实验结果表明,以无编码比值作为DF模型的特征参量,可有效提高变压器故障诊断的准确率。

For the low accuracy of the existing traditional transformer fault diagnosis methods and difficulty in dealing with fault characteristic information effectively,a fault diagnosis method based on deep forest(DF)is proposed.Firstly,considering the correlation between the characteristic gas dissolved in the oil and the fault type,the non-code ratios of the characteristic gas dissolved in the oil are determined as the characteristic parameter of the deep forest model.Then multi-granularity scanning is used to extract more transformer fault feature information.Finally,the diagnosis effect is optimized through cascade forest training.In this case study,the diagnostic effect of the DF model with different feature parameters and different sample sizes is analyzed,compared with multiple traditional diagnostic methods.The experimental results show that using the non-code ratios as the characteristic parameter of the DF model can effectively improve the accuracy of transformer fault diagnosis.