Time series modelling is an influential and successful method for vibration-based applications under data-driven approaches. Because it is a parametric statistical method, one needs to define more details and parameters compare with non-parametric techniques. Time series modelling is generally based on fitting a time series representation to raw vibration measurements and using its statistical characteristics. In vibration-based applications, these characteristics are used for some problems such as system dentification, modal analysis, damage detection, etc. The primary step of time series modelling is to identify an appropriate time series representation is such a way that is should be compatible with the nature of time series data. Although the graphical techniques such as Box-Jenkins methodology are often the initial choices, the model identification via such approaches may be difficult and time-consuming along with some limitations. Therefore, this study proposes an automatic model identification approach by ncorporating the statistical and engineering aspects when vibration time-domain measurements are linear and stationary. In the first step of the proposed approach, it is necessary to perform some data analyses to recognize the nature of vibration time-domain measurements. For the process of model identification, the proposed method relies on numerical evidence based on some information criteria including Akaike’s final prediction error (FPE) and Bayesian information criterion (BIC). The measured vibration responses of an experimental four-story steel structure under ambient excitations are utilized to demonstrate the capability of the proposed method. Results will show that the proposed automatic approach succeeds in identifying the best time series model for linear and stationary time series data and facilitates the process of model identification compared with the Box-Jenkins methodology.

Keywords