一种改进高斯核度量的HEC算法在变压器故障诊断中的应用

Application of Hyper-ellipsoidal Clustering Algorithm Based on Improved Gaussian Kernel Metric in Transformer Fault Diagnosis

针对传统超球型聚类算法难以解决变压器故障诊断问题的特性,使用一种改进的高斯核的超椭球聚类(hyper-ellipsoidal clustering,HEC)算法,并将其解释为寻找体积和密度都紧凑的椭球分簇,该算法能够有效地处理形状为椭球、大小不同和密度不同的分簇。在模拟数据集上的仿真实验表明所提算法在聚类结果和性能上优于K-Means算法、模糊C-Means算法和混合高斯模型期望最大化算法,从而验证了该提算法在处理椭球形或复杂形状数据集聚类时的可行性和有效性;同时将该算法应用在基于变压器油中溶解气体(dissolved gas-in-oil analysis,DGA)的变压器故障诊断中,验证了该方法更高的故障诊断准确度。

In allusion to the problem of traditional hyper sphere clustering algorithm being unable to solve the problem of transformer fault diagnosis, a kind of hyper-ellipsoidal clustering （HEC） algorithm based on improved Gaussian kernel met- ric is used. This HEC algorithm is illustrated as searching for ellipsoid clusters with compact volume and density, which is proved to be effectively handle with clusters of ellipsoid shape with different sizes and densities. Experiment on simulating dataset indicates the proposed HEC algorithm is prior to K-Means algorithm, fuzzy C-Means algorithm and GMM-EM algo- rithm, which verifies feasibility and validity of HEC algorithm in processing problems of ellipsoid dataset or complex-shaped dataset. Application of HEC algorithm in transformer fault diagnosis based on dissolved gas-in-oil also proves higher fault di- agnosis accuracy of this method.