Improved few-shot learning method for transformer fault diagnosis based on approximation space and belief functions

Incomplete and uncertain information is frequently observed in the data analysis processes, which has become one of main challenges for the development of fault diagnosis techniques of transformers. To address few fault cases and deficient monitoring information in diagnostic tasks, this paper provides an improved few-shot learning method based on approximation space and belief functions to accomplish fault diagnosis of transformers. The decision-making table, as an efficient structure to map the weakly correlated attributes, is extracted from transformer cases and maintenance experience. Then the approximation space is used to describe attribute correlations between diagnostic rules and the diagnostic task. We employ the 0.5-approximation set strategy to obtain the diagnostic results when the information is sufficient. Furthermore, we propose a modified basic probability assignment (BPA) calculation method to build belief functions for diagnosis when information is scanty. The modified method is verified capable of improving the decision-making reliability. The overall recognition accuracy of fault diagnosis by our improved few-shot learning algorithm is over 87% which is higher than other four peer methods. This method also shows a potential for good expandability when new diagnostic rules of transformers are discovered.