电力变压器状态监测数据聚类研究

Research on Condition Monitoring Data Clustering of Power Transformer

由于电力变压器油色谱在线监测数据绝大多数为正常数据,只有极少部分为故障或近似故障类数据,这种数据样本的不平衡会导致所训练的变压器故障诊断模型的诊断性能较差。DBSCAN算法是一种基于密度的聚类算法,可以在含有噪声的空间数据库中发现任意形状的簇,适用于含有噪声的状态监测数据聚类。为确保模型训练样本数据的平衡性,文章尝试将DBSCAN算法用于变压器状态监测数据聚类,通过平衡选取聚类后的各类数据样本对变压器故障诊断模型进行训练。工程实例测试结果表明,与随机选取数据样本训练的方法相比,采用该方法所训练的模型具有更优的诊断性能。

Most of the online-monitoring data of power transformer oil chromatography is normal and only a small amount of them is faulted or likely faulted, which leads to poor performance of the diagnosing model being trained. DBSCAN algorithm is a clustering algorithm based on density and can discover clusters of arbitrary shape in spatial databases with noise, which is suitable for the clustering of transformers＇ condition-monitoring data with noise. To ensure the balance of sample data during model training, DBSCAN algorithm is used to cluster transformers＇ condition-monitoring data in the paper, and clustering results are balanced selected to train the diagnosing model. Experimental results show that the method proposed in the paper shows better diagnosing performance compared with that based on randomly selected data samples.