In this paper, the applicability of an auto-regressive model with exogenous inputs (ARX) in the frequency domain to structural health monitoring (SHM) is established. Damage sensitive features that explicitly consider non-linear system input/output relationships are extracted from the ARX model. Furthermore, because of the non-Gaussian nature of the extracted features, Extreme Value Statistics (EVS) is employed to develop a robust damage classifier. EVS provides superior performance to standard statistical methods because the data of interest are in the tails (extremes) of the damage sensitive feature distribution. The suitability of the ARX model, combined with EVS, to non-linear damage detection is demonstrated using vibration data obtained from a laboratory experiment of a three-story building model. It is found that the vibration-based method, while able to discern when damage is present in the structure, is unable to localize the damage to a particular joint. An impedance-based active sensing method using piezoelectric (PZT) material as both an actuator and a sensor is then investigated as an alternative solution to the problem of damage localization. Copyright © 2005 John Wiley & Sons, Ltd.