A comparison of different system identification methods for modal parameter extraction using vibration responses measured from Gisha bridge

In structural monitoring, modal parameters extracted from vibration data are commonly used to gain some information about the condition of bridges. However, even small amount of uncertainty in extracted modal parameters has a considerable erroneous impact on different processes of structural monitoring, including structural model updating and damage detection. Accordingly, in this research effects of different data processing methods and types of vibration tests such as ambient vibration and free vibration, on extracted modal parameters, have been studied. In this regard, four methods including Covariance based Stochastic Subspace Identification (Cov-SSI), Eigensystem Realization Algorithm (ERA), Frequency Domain Decomposition (FDD), and Analytical Mode Decomposition - Hilbert (AMD-Hilbert) have been used to estimate modal parameters. SSI and ERA are parametric methods in time domain in which mathematical bases are similar. FDD and AMD-Hilbert are non-parametric methods which work in frequency and time-frequency domain, respectively. SSI and FDD methods were used for ambient vibration test data and ERA was used for free vibration test records, while AMD-Hilbert method was applied for both free and ambient vibration data. In this article, vibration data of six points were measured from a girder of Gisha Bridge using three Molecular-electronic seismometer sensors, roved in three different setups. One sensor was chosen as reference and its position was fixed among different setups. Data of this sensor were later used for merging different setups results. Therefore, to extract modal parameters multi-setup merging approaches were inevitably used. The measurements were done in vertical direction which leads to identifying vertical bending modes. Ambient vibration responses were measured while the bridge was excited by wind and traffic under the bridge. Free vibration responses were measured after making an impact on the girder.  Two approaches were considered for merging. In the first approach setups were analyzed separately and their final results were combined together and in the second one, merging was done before the process of system identification which eliminates any need to analyze multiple times. A numerical model was also simulated to compare with the field results. Filtering of the recorded data was done before beginning of the system identification process to remove the drift and sudden changes in the signals. Data processing on ambient vibration responses resulted in the first three vertical bending modes which are compatible among all methods, to some extent. In addition, the first two vertical bending modes were identified from free vibration data. Similarity of the mode shapes between different methods were assessed using MAC criterion. Compatible results between these two types of test and numerical model, verifies the results. It is seen that FDD and SSI methods obtained more stable and reliable modal parameters among different setups. Results indicate more modes were identified for ambient vibration data compared to free vibration data. Since, in free response of the structure the first modes are more dominant, lower number of modes could be identified. Considering the non-stationary condition of the conducted vibration tests, the results indicate that the post-processing multi-setup merging approach works better than the pre-processing multi-setup merging approach.