The Effect of Detection Feature Type on Excitations Bred for Active Sensing in Structural Health Monitoring

We have shown that damage detection sensitivity in a structural health monitoring application may be improved by tailoring the excitation via evolutionary algorithms. An improved excitation will be specific to a given application, but prior to this work the role of the features used for damage detection on the input optimization process was unknown. We explore the effect of several detection features on the excitation selected by the optimization. Specifically, we examine three features that are derived from a state-space formulation. In addition, we test detection features based on the residual errors between an autoregressive model and observed data. We show that excitations that have been optimized for a particular feature may still improve sensitivity if another feature is employed, even though optimal classes of solutions for each feature. State-space features that make use of time-dependent correlations tend to select chaotic excitations for improved sensitivity. This contrasts with the single- and multi-tone inputs preferred for the autoregressive features and one version of prediction error. Overall, we provide evidence that damage detection sensitivity can be significantly improved by tailoring the input to match the application and, furthermore, that the optimization routine used to find tailored excitations is general in its application.