Interrogation of the vibration data collected from the sensors embedded throughout the structure without relying on a finite element model of the system for monitoring the health of structural systems has received significant attention in the recent years especially with the current advancements in sensor technology. The data-driven methods explored within this context falls into the realm of statistical pattern recognition field requiring extraction of damage detection features and a statistical decision- making process for identification of damage. Machine learning algorithms provide statistical means for making such decisions. In this study, an unsupervised machine learning approach, one-class support vector machine (OC-SVM), requiring training data only from the undamaged state of the structure is explored for damage detection purposes. The coefficients of the autoregressive (AR) model are extracted as damage sensitive features and used as the required training data. The trained classifier is then used with the data obtained from the same structure at different damage states for classification. Damage detection in the form of recognizing outliers or anomalies not belonging to the target class, is followed by damage localization within the given sensor resolution using statistical means. To this end, Itakura distance measuring the distance between two sets of linear predictor coefficients of the AR processes, is utilized as damage location indicator. Numerical simulations are performed on a truss and a beam structure with several damage scenarios including realistic levels of measurement noise and modeling error. Results show that the proposed approach can successfully detect existence of damage and the statistical measure shows promising performance for further localization of the damaged region.