Time-series analysis of GPS measurements for modeling and estimating bridge movements using Neural Network model: Case study, Tabiat Bridge

One of the parameters that can be used to predict different types' behavior of civil structures is the displacement components, extracted from structures. Bridges and their health monitoring system are important parts of the land transportation system because of their safety and durability. This study aims to evaluate the safety behavior of the Tabiat pavement bridge using a high-rate geodetic monitoring of the Global Positioning System (GPS). Accurate orbital information (SP3) is also used to process the collected observations as well as the Double Difference Bernese software. Firstly, the data are processed kinematically, using the coordinate conversion system, into the local coordinates of the bridge (N, E, and H). At the beginning, the raw data is pre-processed using the Kalman filter to reduce noise. The time series of GPS observations are estimated with 0.0013 m accuracy by applying the Kalman filter with CWPA method. Afterwards, the MA filter, mean time zone and Median filter are used to estimate the semi-static, static and dynamic components of bridge movements, respectively. The results obtained by applying the MA filter indicate that the semi-static component is extracted with 0.005 m accuracy in the North, 0.003 m in the East and 0.01 m in the Height directions. The Median filter also estimates the dynamic component with precision of 0.0007, 0.0006 and 0.001 m in three directions N, E and H, respectively. In this study, considering the sampling rate of observations, the average of 5 minutes of semi-static movement is considered as a static behavior of each point. In order to extract the dominant frequencies of the bridge, the least squares harmonic estimation (LSHE) method is used. Finally the Artificial Neural Network model (ANN) with 5 hidden layers and 5 delays is used to predict bridge deformation based on data mining. The frequencies at points on the bridge are approximately similar, which shows the similarity of the periodogram to the high hardness of the bridge. In addition, the frequencies in both North and East directions, relative to Height direction, indicate the GPS sensitivity to load effects. The prediction model using the artificial neural network is then fitted to dynamic component, with accuracy of approximately 6×10-5m and 4×10-5m to the dynamic and semi-static components of the bridge motion, respectively. Both the least squares harmonic estimation method and the ANN model are considered as suitable techniques for estimating the performance changes of GPS measurements in the frequency and time domains during the monitoring time period. The results indicate that the bridge behavior safety design range the minimum changes using GPS measurements in the time and frequency domains. Besides, the frequency analysis of the bridge movement time series and the neural network model, can be used to detect significant frequency changes and study the bridge performance rigidity, respectively. These results show that high-frequency satellite geodetic data which properly processed can also be useful for measuring the dynamic displacements of tall buildings, cable-stayed bridges, flexible and rigid civil structures. Also they would be the bases of bridge safety monitoring system primary warning.