ABSTRACT: The present work illustrates the recent advances in the damage identification of structural systems in case of “output-only” problems. As safety evaluation and damage assessment of existing engineering structures asks for a deeper structural analysis, usually carried out through the finite element techniques, a numerical model identified in order to properly simulates the actual behaviour of the structure at hand is needed. Consequently identification procedure is an actual engineering concern. To discuss these aspects the experimental campaign conducted in a benchmark three-story steel structure at the Civil Engineering Laboratory of the University of Florence (Italy) laboratory is first reported. The structure has been tested (under ambient loading) starting from the actual configuration and imposing steps of increasingly damage level by partial cuts of flanges on some steel elements (columns). After a neural network based procedures has been proposed and adopted to analyse accelerometer signals recorded during the experimental campaign. Results are discussed and the dynamic properties of the building are evaluated at each damage level. At the end, the ongoing research step aimed to employ genetic algorithms to finite element model identification is discussed. KEY WORDS: Steel-frame structures; Damage identification; Health monitoring; Stochastic Subspace Identification. 1 INTRODUCTION Safety evaluation and damage assessment of existing engineering structures asks for a deeper structural analysis; usually this investigation is carried out through the finite element method techniques: a numerical model of the structure simulates the behaviour of the structure (both in static and dynamic field) and it is used to predict the building response to service and exceptional loads (such as earthquakes). The correct identification of the numerical model is then a fundamental task as the modelling, even if refined from a geometrical point of view, could differ from the real structural behaviour if assumptions on material, constraints, masses and stiffness are not properly evaluated. Variability of these elements can produce results that could be substantially different from the actual behaviour of the structure. To properly identify the numerical model several strategies, or methodologies, are available in the inherent literature. An effective approach to assess the structural behaviour is to evaluate, by performing dynamic experimental tests, actual modal frequencies and corresponding mode shapes. Dynamic tests could be made assuming both forced (vibrodine) and/or environmental (wind or traffic) loads. This step is interesting also with respect to the so-called Structural Health Monitoring (SHM) [5] [1]. The evaluation of these quantities becomes a classical “output-only identification” problem. The identification of the FE model can thus be seen as an optimization strategy where the cost function could be assumed , f.i., the distance between the modal frequencies and mode shape obtained by experimental dynamic tests and those obtained by the numerical model. To address this optimization several approaches have been proposed in literature; among the others quite recently Facchini et al. [6] [7] proposed a neural network approach. The “output-only” procedure could be framed as an emerging topic of both the scientific and technical community as it covers a wide range of practical problems. Effective examples can be considered, for instance, any situation where the need is felt to maintain the structure at hand operative during structural health monitoring investigation: this is the case of bridges which have to remain open to traffic during the test, as well as offshore structures (both small and medium size) [7] where the transport of the equipment needed for the monitoring can be extremely difficult. In structural engineering field the most expeditive (but also, in some respects, the most “brutal”) method to approach the identification problems in case of "output only" systems is the Peak – Picking method (see e.g. Bendat & Piersol 1993 [2]). After the evaluation of the Fourier transform of the recorded signals, the eigenfrequencies of the investigated structure are assessed searching the peaks in the spectrum plot. The eigen- modes can then be determined by comparing the transfer functions of the various recordings with a reference one. The method has the advantage of being relatively simple, however in the case of complex structures it might not be able to provide significant results due to the fact that, of course, it depends on the sensitivity of the operator who has to recognize the peak of auto-spectral density. Despite its disadvantages the popularity of the method is mainly due to its simplicity as the only algorithm that is needed to convert time data to spectra is the Fast Fourier Transform (FFT). In the last decade more advanced methods, even if more computational demanding, have been proposed. Among them it is noteworthy to remember the so-called SSI (Stochastic Subspace Identification) (see e.g. Peeters & De Roeck 1996 [9]). Starting from the consideration that in many cases of Damage identification of output-only systems by means of genetic algorithms Michele Betti 1 , Paolo Biagini 1 , Luca Facchini 1 1 Department of Civil and Environmental Engineering, University of Florence, Via di Santa Marta 3, I-50139 Firenze, Italy email: michele.betti@dicea.unifi.it, paolo.biagini@dicea.unifi.it, luca.facchini@unifi.it Proceedings of the 8th International Conference on Structural Dynamics, EURODYN 2011 Leuven, Belgium, 4-6 July 2011 G. De Roeck, G. Degrande, G. Lombaert, G. M¨ uller (eds.) ISBN 978-90-760-1931-4 2257