With the rapid development of the Internet, particularly social media platforms, user-generated content is being produced at an unprecedented rate, resulting in extremely fast information dissemination. However, this convenience also brings significant challenges, including the rapid spread of false information and misleading content. The rapid spread of rumors poses a serious threat to public awareness, social stability, and even national security. Traditional rumor detection methods mainly focus on propagation structures and textual content, while neglecting model generalization and stability of the model under noise interference. To address these challenges, this study proposes a novel edge-enhanced graph contrastive learning (EGCL) framework, where contrastive learning can learn the connections and differences between similar and dissimilar samples, enhancing generalization and improving accuracy. Specifically, to enhance the stability of contrastive learning, an edge-enhanced module based on a Bayesian uncertainty algorithm is integrated into the graph contrastive learning framework. Experimental results on the public datasets Twitter15, Twitter16, and Weibo show that the proposed method achieves accuracy of 86.7%, 88.2%, and 92.8%, respectively, exhibiting superior performance compared to models from past years.