Transformer Anomaly Detection Based on Multi-Source Information Fusion of Insulation Oil

With the rapid construction and development of smart grids, accurately identifying abnormalities in transformers is of great practical significance for the safe operation of power systems. However, traditional transformer anomaly detection methods make it difficult to effectively detect anomalies inside the transformer from a single source of transformer data. In this paper, we propose a transformer anomaly detection method based on multi-source information fusion of insulating oil, which uses an anomaly detection and diagnosis model with a deep Transform network structure. Specifically, our method contains four modules. The Transformer module is used to receive oil and gas decomposition data in the transformer and uses an attention-based sequence encoder to quickly perform inference knowledge to obtain broader transformer operation time trend data. The Gate Recurrent Unit (GRU) module accepts transformer temperature and bushing temperature data and uses a gating mechanism and a status update mechanism to capture features in the sequence. The Convolutional Neural Networks (CNN) module receives the oil chromatography information at the time of operation of the transformer and uses convolution operations to extract the spatial features in the oil chromatography data. Multilayer Perceptron (MLP) module takes the inputs of the above three modules as input and uses fully connected layers and softmax layers to accurately classify the data. We conducted extensive experiments on real data sets, and the experimental results show that the Accuracy of our proposed method reaches 98.87%, and the F1 reaches 99.40%, surpassing all baseline classifiers, and the F1 exceeds the state-of-the-art method by 3.77%. Furthermore, we explore the impact of different models and features on detection results.