Dynamic decomposition and multilevel feature fusion for dissolved gas in oil anomaly detection

The non-stationarity and dynamic change problems existing in the time series data lead to the limited accuracy in outlier identification. Therefore, this paper proposes an outlier identification method based on dynamic decomposition and Multi-feature Fusion (DDMMF). Firstly, the dynamic decomposition method based on deep residual separation effectively captures the correlations and co-variation patterns among different gas components. Secondly, a multi-scale feature fusion strategy is introduced to alleviate the problem of local information loss caused by the limitation of internal potential space and enhance the accuracy of data reconstruction. And the reconstruction error is calculated to achieve the effective detection of outliers. The experiment uses 1000 kV and 110 kV transformer dissolved gas data to verify the model’s validity and generalization across voltage levels. Experimental results show that the outlier recognition accuracy of DDMMF reaches 0.947, outperforming other comparison methods.