An efficient procedure is developed for health monitoring of bridge by using ambient vibration due to traffic loads. The main algorithms consist of the random decrement method incorporating band-pass filters for estimation of the free vibration signals, the cross spectral density function method for identification of the modal parameters, and the neural network technique for estimation of the element-level stiffness changes. An experimental study is carried out on a scaled bridge model with a composite section subjected to various moving vehicle loadings. Vertical accelerations are measured at several locations on the girder. The estimated frequencies and mode shapes are found to be well-compared with those obtained from the impact test. A finite element model is constructed, which consists of beam elements for the girder and rotational spring elements at the ends. The baseline values of the stiffness properties of the elements are estimated using the neural network technique. Then the stiffness parameters of the damaged cases are evaluated based on the corresponding modal properties. The estimated stiffness changes using the neural networks are found to be very good for the case with the simulated data. However further study is needed for verification based on the real measurement data.