UC San Diego
Evolutionary algorithms, chaotic excitations, and structural health monitor : on global search methods for improved damage detection via tailored inputs
- Author(s): Olson, Colin C.
- et al.
In a vibration-based damage detection paradigm, a structural health monitoring (SHM) system is designed to acquire data from which damage-sensitive features will be observed in order to classify the damage state of the structure without providing false positive indications of damage. An optimization routine is introduced that exploits the concept of SHM as a pattern recognition problem for which there is a space of solutions that can be searched by global optimization algorithms to improve damage detection sensitivity. The optimization procedure is general in that it can be used to improve a number of elements that are particular to the problem of damage detection including: the excitation, method of data conditioning, detection features, and statistical metrics of comparison. In particular, this work focuses on tailoring excitations for global, active sensing SHM applications via evolutionary algorithms. Indications of damage in the response are shown to be more detectable in both computational and experimental platforms using excitations that have been tailored to the application via the optimization routine. Using the routine, a class of excitations that significantly enhance damage detection relative to untailored inputs is discovered and shown to do so because of a change in the state-space representation of structural response. When state-space features are employed, the relative power of structural response frequencies is shown, in the most basic case, to significantly affect the 2-torus representation of the response. The routine is also used to show that sensitivity is significantly affected by the employed detection feature. In particular, the introduction of temporal information in the feature formulation dictates that improved detection occurs for a class of chaotic excitations. In addition, a method is introduced that allows tailored inputs to be generated for more complicated structures by training on a simple infinite impulse response filter. Design of the filter only requires identification of two structural resonant frequencies. A 6% reduction in bolt stiffness in an experimental structure is clearly classified as damaged using state-space features and excitations that have been trained on the filter. The reduction in stiffness is not visible using traditional modal methods