Earthquake has been the most threatening disaster to civil structures. The damage of infrastructures not only causes the loss in economic activities but the life and properties of people. For that reason, research on applying instrumentation on structures to monitor their characteristic during and after earthquakes is always an important issue. By the development of system identification in monitoring the vibration trait of buildings and bridges, some theories and methods has gradually become matured in the last few decades. In order to record the earthquake data in Taiwan, the Central Weather Bureau (CWB) of the Ministry of Transportation and Communication (MOTC) has instrumented sensors in free field, buildings and bridges among Taiwan to record their response during major earthquakes. For the part of bridges, sixteen bridges have been chosen as the demonstration examples and a safety monitoring and seismic evaluation system has been established. The database collected by the system can offer abundant information for researchers. The practical situation of bridges can be reflected by the data and a proper modification of design code can be determined. Traditionally, the identification methods were developed under frequency domain. However, two close frequency modes can not be effectively separated by frequency-domain-based method when noise is contained in the measured data. To solve this problem, the discrete-time domainbased identification technique has been applied to civil engineering in the last two decades. In most methods, structures are considered time-invariant. Namely, parameters of structures are assumed to be constant during the whole time history. However, structures might be damaged or with nonlinear behavior during earthquake. To this goal, neural networks, proved to have outstanding performance in complex problems, was integrated into identification systems. The adaptability and fault tolerance of neural networks have made them good candidates in dealing with data of uncertainty and incompleteness and identification of nonlinear systems under major earthquakes may be implemented by neural networks (Adeli et al. 1995) (Masri et al. 1992) (Masri et al. 1993). A bridge health monitoring system based on neural network technology is proposed in this paper. In order to identify the nonlinear behavior of structures, a NARX system is trained from data collected in major earthquakes. The relationship between the input and output channel can be reflected by the weighting of the neural network and the fundamental period of the structure can then be derived. By applying the system to bridges, the multi-support characteristic can be analyzed and the combination of specific frequencies causing resonant phenomenon can also be obtained. The result would be an important basis for verifying the source of damage behavior on structures. To demonstrate the performance of the proposed system, a bridge in the second southern freeway in Taiwan is used. By data collected from two large ground excitations, the NARX system with a structure of two input nodes and one output node is established to evaluate the property of the bridge. The input channels are the signals from the pile-cap and the output is the response of the web on the middle span. Analytical results of different methods including transfer function, ARX model are also compared with the proposed neural-network-based system to evaluate their efficiency in health monitoring. The result has shown that besides identifying the fundamental frequency of structure, the proposed neural-network-based system can also be successfully applied in bridge health monitoring after major earthquakes. The combination of specific frequencies causing resonant phenomenon is clearly shown in the kernel transformation diagrams and damage on structures could be alleviated. More information can be studied from the complex high-order kernel transformation. The capability of the NARX system in dealing with nonlinear structure would be another research focus. By the proposed method, bridges with nonlinear bearing such as lead rubber bearing or viscoelastic dampers can be precisely monitored during major earthquakes. The behaviour of these elements under specific time history can be used to assess the performance of these equipments and the bridge design code can be revised basing on the practical condition of bridge structures.