近邻成分分析和k近邻学习融合的变压器不平衡样本故障诊断

Transformer Fault Diagnosis with Unbalanced Samples Based on Neighborhood Component Analysis and k-Nearest Neighbors

在基于机器学习的电力变压器故障诊断方法中,各故障类别间案例数量不平衡会导致诊断准确率降低。为了提升电力变压器故障诊断模型的准确率及运行效率,构建了融合引入修正因子的近邻成分分析和k近邻学习的故障诊断模型。首先,通过对近邻成分分析算法(neighborhood component analysis,NCA)目标函数引入修正因子减少样本不均衡对模型训练的影响,结合油色谱故障数据通过关联规则得到样本参量相关性量化矩阵,作为NCA算法训练度量矩阵的初值;然后,利用训练得到的度量矩阵对k近邻(k-nearest neighbors,k NN)分类器的输入数据结果进行映射变换,使同类型样本间的距离减小,进而使k NN分类性能提高;最后,用贝叶斯优化算法对模型进行超参数调优,获得能使测试集准确率最高的模型参数集。以变压器故障案例库为对象的算例分析结果表明,提出的模型与传统的机器学习诊断模型相比,用时节省了近一半,且所提模型对少数样本类的诊断准确率相比于其它模型提升了至少15%。论文研究可为电力变压器的故障诊断提供参考。

In transformer fault diagnosis models based on machine learning,the imbalance of class distribution among fault categories will lead to low diagnosis accuracy.In order to improve the accuracy and the efficiency of the models,a fault diagnosis model based on k-nearest neighbors and neighborhood component analysis with correction factors was proposed.Firstly a correction factor to NCA algorithm’s objective function,was introduced to reduce the impact of sample imbalance on model training.Also,association rules to obtain a parameter correlation quantization matrix according to the fault data,was used as the initial value of NCA algorithm’s training metric matrix.Secondly,the result matrix of NCA was used to map the input data of k NN classifier,which could reduce the distance between samples of the same class and improve the classification performance of the classifier.Finally,the Bayesian optimization algorithm was applied to tune the hyper-parameters of the model to obtain the model parameter set that maximizes the accuracy of test set.The practical value of our model was verified through the data from dataset.The experimental results show that,compared with the traditional machine learning diagnosis models,the proposed method can save nearly half of time,and the diagnostic accuracy of minority classes is improved by at least 15%.The research can provide references for the fault diagnosis of power transformers.