In the process of on-line monitoring transformers by the dissolved gas analysis（DGA）

the data are oftendistorted due to the problems of monitoring devices

which seriously affects the effectiveness and practicability of theentire on-line monitoring system. The traditional threshold methods cannot detect the abnormality of transformer andmonitoring device flexibly and timely. To deal with this issue

a DGA data outlier detection method based on slidingwindows and multivariate Gaussian distribution is proposed. First

exponentially weighted moving average（EWMA）and sliding window were used to obtain the deviation data sets between the predicted values and the real values. Then

an anomaly detecting model was established by multivariate Gaussian distribution. The measured data of oil gas of onetransformer in substation were used to validate the method. The results reveal that the method can detect not only theabnormality of transformer

but also the data distortion of monitoring device in real time. And the impact of data muta-tion and noise can be eliminated. Overall

the method has practical value for the anomalous state detection.