A damage identification method based on the vector auto-regression (VAR) time series is proposed. First, the VAR model is trained by the measured acceleration time-histories, and a new vector is extracted from the diagonal elements of the VAR coefficient matrices as a damage sensitive vector. The Mahalanobis distance of the new vector is defined as the damage feature. Then, the area under a receiver operating characteristic (ROC) curve is used to detect and localize the damage using statistics pattern recognition methodology, and the Bhattacharyya distance (BD) is implemented to quantify the magnitude of damage. Results from a simulated model and a lab-scale frame structure show that damage location and severity can be successfully identified and a reasonable amount of noise can be canceled.