The operation of power transformers involves significantly fewer fault data compared to normal data, resulting in a severe data imbalance issue. Additionally, the complex coupling relationships among the monitored variables make the modeling of condition assessment tasks challenging and lead to low evaluation accuracy. Aiming at the related problems, a power transformer health condition evaluation method based on double input residual graph convolutional network is proposed. First, SMOTE Tomek mixed sampling algorithm was used to pre-process the unbalanced data of the training data, which solved the problem of insufficient fault data and difficult classification. Then, a multi-metric fusion graph construction method is proposed to learn the correlation between variables from multiple variables and construct the graph structure data. Finally, a double input residual graph convolutional network(DI-ResGCN) based on the ChebyNet is proposed, feature extraction is carried out on the constructed graph structure data, and feature fusion is carried out through the selfattention mechanism to obtain the transformer health evaluation results. Experiments were carried out on a dataset of dissolved gases in oil and oil test collected by a real power transformer, and experimental results show that the proposed method has a state assessment accuracy of 94.7% and F1 score of 0.942, outperforming other deep learning methods and exhibiting the best evaluation performance.