Dissolved Gas Analysis (DGA) plays a critical role in the fault detection of oil-immersed power transformers. However, the substantial imbalance between fault and normal classes in DGA datasets poses a significant challenge, often leading to biased models with suboptimal diagnostic performance. To address this issue, we propose GCS, a graph-augmented semi-supervised contrastive learning approach. GCS constructs a \( K \) NN graph to capture the complex relationships among samples, which is then integrated into a semi-supervised contrastive learning framework. This integration allows for the refinement of DGA sample embeddings by leveraging label information. Furthermore, GCS focuses on expanding the representation of minority class samples within the embedding space, ensuring comprehensive learning of their characteristics. Experimental results on two datasets demonstrate the superior diagnostic performance of GCS in scenarios of class imbalance, confirming its effectiveness in enhancing the reliability of fault diagnosis in power transformers.