变压器不平衡样本故障诊断的过采样和代价敏感算法

Oversampling and Cost-sensitive Algorithm for Transformer Fault Diagnosis with Unbalanced Samples

基于神经网络的变压器故障诊断方法是评估变压器状态的重要方法,然而该方法在处理各状态类别间样本数量不平衡的数据集时,各类型状态识别效果差距较大,识别结果更多偏向多数类样本。文中基于过采样方法和代价敏感算法,针对最大不平衡度为266∶19的油色谱数据集,构建了一种用于变压器故障诊断的BPNN模型。首先,基于SMOTE算法对样本数据集进行有选择的扩充,该算法结合了近邻分析和线性插值的思想,减少了样本扩充所导致的模型训练的过拟合现象。然后,构建五层神经网络,并引入Focal Loss函数取代交叉熵函数作为本模型的代价损失函数,从而更关注于少数类样本的识别和区分。实验结果表明,文中模型相比于原始BPNN模型在总体准确率上提升了6.48%,各少数类样本类别的F1分数分别提高了25.7%、11.4%、3%、26.1%、1.8%、15.3%和33.3%,并且算法收敛更快。在和传统机器学习方法的对比中,文中模型比基于KNN算法和随机森林算法模型的整体准确率分别提高了16.53%和7.98%。

Fault diagnosis method of transformer based on neural network is an important method to assess the states of transformer.However,when processing data sets with unbalanced sample numbers among various fault categories by the method,the effect of different types of state recognition is large and the fault recognition effects are more biased to the majority samples.For the oil chromatography data set with a maximum imbalance of 266∶19,a kind of BPNN model for transformer fault diagnosis based on oversampling method and cost-sensitive algorithm is constructed.First,the sample data set is selectively expanded based on the SMOTE algorithm,which combines the ideas of nearest neighbor analysis and linear interpolation to reduce the over fitting of model training caused by the sample expansion.Then,a five-layer neural network is constructed,and the Focal Loss function is introduced to replace the cross-entropy function as the cost loss function of this model,thereby paying more attention to the identification and differentiation of minority samples.The experimental result shows that the overall accuracy of the model in this paper is improved by 6.48%compared to the original BPNN model,and the F1 scores of each minority sample category are improved by 25.7%,11.4%,3%,26.1%,1.8%,15.3%and 33.3%,with faster algorithm convergence speed.In comparison with traditional machine learning method,the overall accuracy of the model in this paper is improved by 16.53%and 7.98%compared to the model based on KNN algorithm and random forest algorithm.