LMS‐based structural health monitoring of a non‐linear rocking structure

A structure's health or level of damage can be monitored by identifying changes in structural or modal parameters. This research directly identifies changes in structural stiffness due to modelling error or damage for a post‐tensioned pre‐cast reinforced concrete frame building with rocking beam column connections and added damping and stiffness (ADAS) elements. A structural health monitoring (SHM) method based on adaptive least mean squares (LMS) filtering theory is presented that identifies changes from a simple baseline model of the structure. This method is able to track changes in the stiffness matrix, identifying when the building is (1) rocking, (2) moving in a hybrid rocking–elastic regime, or (3) responding linearly. Results are compared for two different LMS‐based SHM methods using an L ₂ error norm metric. In addition, two baseline models of the structure, one using tangential stiffness and the second a more accurate bi‐linear stiffness model, are employed. The impact of baseline model complexity is then delineated. The LMS‐based methods are able to track the non‐linearity of the system to within 15% using this metric, with the error due primarily to filter convergence rates as the structural response changes regimes while undergoing the El Centro ground motion record. The use of a bi‐linear baseline model for the SHM problem is shown to result in error metrics that are at least 50% lower than those for the tangential baseline model. Errors of 5–15% with this L ₂ error norm are fairly stringent compared to the greater than 2 × changes in stiffness undergone by the structure, however, in practice the usefulness of the results is dependent on the resolution required by the user. The impact of sampling rate is shown to be negligible over the range of 200–1000Hz, along with the choice of LMS‐based SHM method. The choice of baseline model and its level of knowledge about the actual structure is seen to be the dominant factor in achieving good results. The methods presented require 2.8–14.0 Mcycles of computation and therefore could easily be implemented in real time. Copyright © 2005 John Wiley & Sons, Ltd.