Power Transformer Fault Diagnosis Considering Data Imbalance Based on Counterfactual Graph Convolutional Network

Improving the accuracy of transformer dissolved gas analysis plays a vital role in electric power systems. Due to the scarcity of high-quality fault data, prevailing methods for diagnosing power transformer faults rely on a copious amount of health data, which considerably exceeds the quantity of fault data, in an effort to enhance diagnostic accuracy. However, this approach encounters a significant challenge due to the highly imbalanced nature of the data (i.e., a predominant proportion of health data), which causes minimal variation in the gradient of classifiers during the training phase under conventional loss functions. Simply, the classifier only needs to predict health to achieve high accuracy rates. In this paper, we propose a novel method for diagnosis based on the graph convolutional network combining counterfactual analysis. Specifically, we employ the structural causal model to dissociate the direct causal effect of sample quantity on predictions from the total causal effect. We evaluate our proposed method on the data set comprising 10113 health samples and 1127 fault samples. The experimental results demonstrate that our method effectively mitigates the detrimental impact of data imbalance and utilizes health samples to enhance the accuracy of transformer fault diagnosis, particularly in scenarios where fault data is scarce.