Transformer fault diagnosis technology based on AdaBoost enhanced transferred convolutional neural network

Transformer fault diagnosis is an important means to ensure the stable operation of power systems, and Dissolved Gas Analysis (DGA) is a widely used method. However, existing intelligent diagnostic methods have problems such as insufficient expression of gas features, poor generalization of classification models, and lack of specialized deep optimization. On this basis, the authors proposed a transformer fault diagnosis technology based on AdaBoost enhanced transferred convolutional neural network, and constructed a transformer fault diagnosis scheme of feature engineering, classification models, specialized deep optimization. Firstly, the original five-dimensional gas features were expanded into new twenty-one-dimensional gas features to address the problem of insufficient expression of gas features. Factor analysis method was used to extract key information from twenty-one-dimensional gas features, and twelve-dimensional common factors were transformed to create the high-quality DGA data set. Secondly, the ensemble learning idea was referenced to construct ensemble classification models, and the best performing AdaBoost was obtained to solve the problem of poor generalization of single classification models. Finally, the transfer learning idea was referenced, the convolutional neural network was used as the basic classification model for AdaBoost, learning parameters were passed, and AdaBoost-TCNN with superior and stable classification effect was constructed to solve the problem of lack of specialized deep optimization. Test results on both private and public data sets show that, ensemble classification models achieved higher classification accuracy than single classification models, with significant advantages in Precision, Recall, and F1-score compared to single classification models. Among them, AdaBoost-TCNN achieved the highest classification accuracy of 97.33% and 95.35%, respectively, and obtained the most stable and significant Precision, Recall, and F1-score, all of which were superior to the ensemble classification models in the latest research. In addition, AdaBoost-TCNN also demonstrated a focus on key fault types. Therefore, the necessity of conducting this study, as well as the feasibility, practicality, and superiority of the technology proposed in this paper, have been fully demonstrated. © 2024 Elsevier Ltd