The online monitoring of power transformer is an important guarantee of operation situation. The traditional method is based on the threshold in guidelines or related documents of state grid

and the anomalous situation can not be detected in time. Aiming at solving these problems

we proposed a fast anomalous state detection method of power transformers based on sliding windows and clustering. First

the time series and sliding windows were used to extract features of multi-dimensional state data. The anomalous time and type were recorded in order to establish the detecting model of anomalous data. Then

an anomaly detecting model were established by the unsupervised k-means clustering method. This model was used for detection of the multi-dimensional online monitoring data so that the anomalous time and type were detected. The measured data of oil gas of one transformer in substation were used to validate the method. The results reveal that the anomalous trends of online monitoring data stream can be detected in time

which are caused by the changing of operation situations. And the disturbance of few noise data and abrupt change point can be eliminated. Overall

the method is valid for the anomalous state detection.