Damage detection of a highway bridge under severe temperature changes using extended Kalman filter trained neural network

Detecting structural damage in civil engineering structures has become an increasingly viable option for efficient maintenance and management of infrastructures. Vibration-based damage detection methods have been widely used for structural health monitoring. However, those methods may not be effective when modal properties have significant variance under environmental effects, especially severe temperature changes. In this paper, an extended Kalman filter-based artificial neural network (EKFNN) method is developed to eliminate the temperature effects and detect damage for structures equipped with long-term monitoring systems. Based on the vibration acceleration and temperature data obtained from an in-service highway bridge located in Connecticut, United States, the correlations between natural frequencies and temperature are analyzed to select proper input variables for the neural network model. Weights of the neural network are estimated by extended Kalman filter, which is also used to derive the confidence intervals of the natural frequencies to detect the damage. A year-long monitoring data are fed into the developed neural network for the training purpose. To assess the changes of natural frequencies in real structural damages, structural damage scenarios are simulated in the finite element model. Numerical testing results show that the temperature-induced changes in natural frequencies have been considered prior to the establishment of the threshold in the damage warning system, and the simulated damages have been successfully captured. The advantages of EKFNN method are presented through comparing with benchmark multiple linear regressions method, showing the potential of this method for structural health monitoring of highway bridge structures.