Output-only structural health monitoring in changing environmental conditions by means of nonlinear system identification

Abstract

Structural health monitoring relies on the repeated observation of damage-sensitive features such as strains or natural frequencies. A major problem is that regular changes in temperature, relative humidity, operational loading, and so on also influence those features. This influence is in general nonlinear and it affects different features in a different way. In this article, an improved technique based on kernel principal component analysis is developed for eliminating environmental and operational influences. It enables the estimation of a general nonlinear system model in a computationally very efficient way. The technique is output-only, which implies that only the damage-sensitive features need to be measured, not the environmental parameters. The nonlinear output-only model is identified by fitting it to the damage-sensitive features during a phase in which the structure is undamaged. Afterwards, the structure is monitored by comparing the model predictions with the observed features. The technique is validated with natural frequency data from a three-span prestressed concrete bridge, which was progressively damaged at the end of a one-year monitoring period. It is demonstrated that capturing the regular variations of the features requires a nonlinear model. Monitoring the misfit between the predictions made with this model and the observed data allows a very clear discrimination between validation data in undamaged and damaged conditions.