Defect Pattern Recognition of Transformers Based on the method of vectorization of parameters

It is important for the maintenance department of Power Company to recognize the defect pattern of transformers. The traditional way of defect pattern recognition is mainly dependent on the expert-experience and rarely driven by data analysis. Therefore, this paper proposes a defect pattern recognition method based on vectorization of condition parameter and Huffman tree along with the neural network principle. With this method, the condition parameters are transferred into a certain length of vector as the input in order to reduce the dimensionality. For the output layer, the Huffman-tree of defect pattern has been built to reduce the calculation task during backpropagation. By accumulating recent 3 years defect data of transformers, this model has been trained and it is shown that the accuracy of recognition reaches to 94.3%. Finally, the cosine correlation between condition indices and the internal connection have been obtained by applying the trained parameter vector.