Health monitoring of large structures is inherently difficult due to the relatively small number of available sensors/measurements that can be made within budgetary constraints. To accurately detect the presence of damage in a structure requires a reliable model, or at least a good representation of the structure prior to damage. Approaches to detecting and localizing damage are predominantly based on either frequency changes or transient responses. Transient or closed-loop responses are available more readily during operation and appear to be more suitable for online damage detection than approaches based on frequency changes. In order to detect damage in a large structure, the structural characteristics such as mass matrices and stiffness matrices need to be estimated. This paper utilizes an implementation of the unscented Kalman filter in square-root form to estimate changes to the system mass/stiffness. The damage detection problem is solved online by updating the structural parameter estimates using a limited amount of measurement data. Example results are presented for spring–mass, beam and truss structures where the only measurements are accelerometer data from a limited number of nodes. The numerical results show that the approach is capable of detecting changes in the structure from the outputs online.