Abstract

This paper proposes a new framework, named BDD, which bridges Duval’s method with a deep neural network (DNN) approach for power transformer fault diagnosis using dissolved gas analysis (DGA). The proposed BDD consists of the following three key points. First, to overcome an important issue that most DGA data found in real-world industrial settings is unlabeled, Duval’s method is newly used to provide knowledge, which is called pseudo-labeling information, to a DNN for unlabeled DGA data. Second, motivated by the fact that the pseudo-labeled data does not always declare correct answers, a DNN architecture with an auxiliary regularization task is newly proposed, which is somewhat robust to the noisy labeled data. Last, a parameter transfer learning approach is applied to evolve the pre-trained DNN model, which is trained from a large amount of pseudo-labeled source DGA data, for diagnosing the sparse labeled target DGA data. Four case studies are executed through the use of KEPCO's massive unlabeled DGA data and IEC TC 10′s sparse labeled DGA data: (i) a comparison with the existing methods, (ii) examination of the effectiveness of parameter freezing via feature space investigation, (iii) studying the robustness of the regularization task under noisy labeled DGA, and (iv) probing the hyperparameter effects. Moreover, to strengthen the proposed model’s effectiveness, the last fifth case study performs a comparison with the existing methods for KEPCO's sparse labeled data instead of IEC TC 10 data. We confirm that the proposed BDD method outperforms existing methods, thanks to the Duval method’s weak supervision, the regularization task, and parameter transfer.