Anomalous State Detection of Dissolved Gases in Transformer Oil Based on the Canopy Hyper Sphere Model

Abstract

The dissolved gas in oil is one of major state parameters of power transformers. The anomaly must be recognized before fault diagnose. However, data fluctuation and missing may cause anomaly recognition methods inapplicable. In this paper, a new method of abnormal state rapid identification of transformer is presented based on the Canopy model. The Canopy algorithm can determine the cluster number and cluster center position in the case of unknown state class, and has the advantages of small amount of calculation and fast convergence. This paper analyses the error of gases in oil detecting data and proposes the outlier recognition method based on the sliding window. Evaluation of data quality by the introducing fluctuation coefficient and variable weight high dimensional space is established. In the variable weight high dimensional space, the improved Canopy model is used to distinguish the state, and the abnormal event is used to identify the abnormal state. Compared with K-Means, the method improves the boundary data classification effect and reduces the computational complexity. With the variation tendency judgment, the anomaly state can be recognized. By testing with a not exceed standard practical cases, the method effectively recognized the overheat defect. And the method also does well in the threshold false alarm cases that caused by interference or poor data quality.