Physics-based approaches to structural health monitoring with vibration-based data rely heavily on structural identification, or finite element model updating, to develop both baseline reference and damage models for diagnostics and severity estimation. However, practical challenges associated with the limited number of measurement sensors and finite number of measured mode shapes produce a difficult, nonlinear inverse eigenvalue problem often plagued by multiple local minima, potentially non-smooth objective functions, and a large search space. Genetic algorithms can be adopted to address many of these challenges by providing a bounded and constrained global optimization technique capable of yielding globally minimized objective functions for structural identification. However, the drawback of genetic algorithms is the necessity to evaluate a very large number of candidate models over many iterations that converge slowly. For large finite element models, this can yield solution times that are impractical for implementation. Fortunately, genetic algorithms are highly parallelizable and can experience dramatic improvements in computational speed using parallel computing clusters. This paper presents a strategy and framework for implementing parallel genetic algorithms for structural identification of civil structures using commodity microprocessors, including low-cost single board computers. Design of the cluster is detailed and experiences with application to a high fidelity finite element model of a highway bridge are presented.
doi: 10.12783/SHM2015/112