The large integration of modular multilevel converters (MMC) has introduced stability issues. The impedance-based stability analysis method is widely adopted, where the impedance model can be directly achieved at the terminals through nonintrusive measurement, which facilitates the black-box stability analysis of the MMC-grid interaction system. Yet, due to the limited impedance data amount in the practical application of online impedance identification, the accuracy of the identified model stability analysis cannot be guaranteed with existing methods in variable operating point scenarios. This article proposes a transfer learning based online impedance identification for MMC to address this research gap. The two-phase online impedance identification method is developed where the physical model of MMC in the offline phase is utilized to facilitate the online impedance identification. The proposed method can significantly reduce the data amount requirement in online impedance identification and achieve online stability analysis of the MMC system. The case studies confirm the effectiveness of the proposed method.