A major difficulty of structural identification is due to the large number of unknowns and its induced ill condition. Substructural identification approach opens a new angle for the identification of large-scale structures because it offers the flexibility to isolate some critical substructures for identification. The key is to estimate the boundary force using response measurements. However, after 2 decades of development, it has been found that this approach worsens the identifiability of the inverse problem. In this paper, we propose a new real-time substructural identification approach to resolve the ill-posed problem. Its major crux is to model the boundary force as modulated filtered white noise, which can be viewed as a continuity condition. As a result, the boundary force is estimated not only through the response measurement at the same time step but also the boundary force estimation of previous time steps. This drastically enhances the computational condition of the problem. Furthermore, the proposed method does not require stationarity of the response. Examples using a hundred-story shear building under different stationarity scenarios are used to demonstrate the performance of the proposed method.