基于NNTR-SMOTE与GA-XGBoost的变压器故障诊断方法研究

Transformer fault diagnosis method based on NNTR-SMOTE and GA-XGBoost

针对变压器故障诊断中故障样本数量少且分布不均衡导致诊断率低的问题,提出了一种基于最近邻三角区域合成少数类过采样(NNTR-SMOTE)与利用遗传算法(GA)优化极端梯度提升(XGBoost)模型的变压器故障诊断方法。首先,将采集到的变压器故障样本数据进行标准化处理,使用NNTR-SMOTE方法得到平衡数据;其次,采用无编码比值法构造油中溶解气体的特征,得到特征数据集并对特征数据集采用多维尺度分析(MDS)方法进行特征融合;最后,利用GA对XGBoost模型的参数进行优化,构建变压器故障诊断模型。试验结果表明:基于NNTR-SMOTE与GA-XGBoost的变压器故障诊断方法诊断准确率达95.97%,不仅解决了诊断模型对多数类的偏向问题,还将模型的诊断精度进一步提高,适用于变压器非均衡数据集的多分类故障诊断。

To address the low accuracy of transformer fault diagnosis caused by the insufficient number and uneven distribution of fault samples,a transformer fault diagnosis method based on nearest neighbour triangle regions synthetic minority oversampling technique(NNTR-SMOTE)and genetic algorithm optimized extreme gradient boosting(GA-XGBoost)is proposed.Firstly,the transformer fault sample data are collected and standardized,and then balanced data are obtained by NNTR-SMOTE.Secondly,the feature data of dissolved gas are categorized by non-coding ratio method,and then fused by multi-dimensional scaling(MDS)method.Finally,a new transformer fault diagnosis model is constructed based on the XGBoost model optimized by GA.The experimental results show that the diagnostic accuracy of the proposed diagnosis method based on NNTR-SMOTE and GA-XGBoost reaches as high as 95.97%.This method not only solves the bias towards the majority class during diagnosis modelling,but also improves the diagnostic accuracy of the model,making it suitable for multi-classification fault diagnosis for transformers with unbalanced data.