To identify the model structure parameters in shaking table tests from seismic response, especially from timevarying response records, this paper presents a new methodology by combining the online recursive Adaptive Forgetting through Multiple Models(AFMM) and offline Auto-Regression with eXogenous variables(ARX) model. First, the AFMM is employed to detect whether the response of model structure is time-invariant or time-varying when subjected to strong motions. Second, if the response is time-invariant, the modal parameters are identifi ed from the entire response record, such as the acceleration time-history using the ARX model. If the response is time-varying, the acceleration record is divided into three segments according to the accurate time-varying points detected by AFMM, and parameters are identifi ed by only using the tail segment data, which is time-invariant and suited for analysis by the ARX model. Finally, the changes in dynamic properties due to various strong motions are obtained using the presented methodology. The feasibility and advantages of the method are demonstrated by identifying the modal parameters of a 12-story reinforced concrete(RC) frame structure in a shaking table test. To identify the model structure parameters in shaking table tests from seismic response, especially from timevarying response records, this paper presents a new methodology by combining the online recursive Adaptive Forgetting through Multiple Models (AFMM) and offline Auto-Regression with eXogenous variables (ARX) First, the AFMM is employed to detect whether the response of the model structure is time-invariant or time-varying when subjected to strong motions. Second, if the response is time-invariant, the modal parameters are identifi ed from the entire response record, such as the acceleration time-history using the ARX model. If the response is time-varying, the acceleration record is divided into three segments according to the accurate time-varying points detected by AFMM, and parameters are identifi ed by only using the tail segment data, which is time-invariant and suited for analysis by the ARX model. Finally, the changes in dynamic properties due to various strong motions a The feasibility and advantages of the method are demonstrated by identifying the modal parameters of a 12-story reinforced concrete (RC) frame structure in a shaking table test.