Abstract: A new method is developed for identifying structural damages at the element level by using time-domain response data at a few points caused by random loadings. In the study, a time series model with a fitting order is first constructed using the time domain response data with measurement noise. A sensitivity matrix consisting of the first differential of the autoregressive coefficients of the time series models with respect to the stiffness of the structural elements is then obtained based on simulated time domain response data. Finally, the locations and severities of the damage can be easily obtained by solving for the damage vector whose components are the damage degrees of the structural elements. The efficiency and capability of the proposed method are demonstrated by applying the method to a cantilever beam with damage at one or two elements. Numerical simulations show that the use of one sensor acceleration history data with 1% and 2% measurement noises is capable of identifying damages at a single element and multiple elements efficiently and the increase in numbers of sensors is certainly helpful for improving the diagnosis correction ratio.