The stability problems caused by converter interaction challenge the safe operation of DC microgrids. Online stability monitoring methods need to be developed urgently to prevent instability risk for DC microgrids. The existing measurement-based methods are invasive and need widely-deployed measurement units, which is difficult to realize. In this paper, a novel semi-supervised data-driven stability monitoring method based on empirical mode decomposition (EMD) and semi-generative adversarial network(SGAN) is first proposed in DC microgrids. The instability oscillation information can be extracted from complex DC bus voltage by the EMD algorithm, and the semi-supervised learning framework based on SGAN completes the training process with easily accessible unlabeled data and much less labeled data. The proposed method is non-invasive and significantly reduces reliance on labeled data, in contrast to existing methods. Subsequently, experimental tests on a hardware-in-loop(HIL) test platform confirm the effectiveness and high performance of the method, as evidenced by the construction of a DC microgrid model.