Accurate anomaly localization can ensure the continuity and reliability of power supply in DC microgrid. However, it is still a challenge to improve the robustness of anomaly localization methods under the complexity and dynamics of DC microgrids. To address this issue, an adversarial autoencoder method is proposed to improve the generalization of anomaly localization and sensitivity for anomaly data. Firstly, the physical-cyber model of microgrid system is established, and the attack model under false data injection attack is analyzed. Then, an adversarial autoencoder is proposed to accurately locate anomalies, enhancing its robustness through noise and reducing false positives with a discriminator. Consequently, the DC microgrid anomaly localization simulation platform was constructed to obtain the dataset for training the localization model. Extensive experiments are conducted to validate the proposed method, which can improve the recall rate by up to 33.3% and reduce the false positive rate by 8.2%.