Fully automated OMA : an opportunity for smart SHM systems

Advances in dynamic identification procedures and optimization of hardware performances play a relevant role in the development of Structural Health Monitoring in hazardous areas. Several worldwide applications are reported in the literature and several methods able to assess the health state of a structure exist, some of which are based on tracking the modal characteristics of the structure during service life and especially after damage due to exceptional loads. The most relevant drawbacks of such methods, however, is represented by the need of a user intervention during the modal parameter identification processes, that does not fit requirements of SHM systems. In this paper, an approach for automated modal parameters identification and tracking is described: the algorithm has been integrated in a fully automated SHM system and is based on a consolidated technique of operational modal analysis, the Frequency Domain Decomposition. The algorithm has been implemented into a specific software package developed in LabView 8 environment and it is still submitted to extensive tests. Some results obtained since its integration in the SHM system of the School of Engineering Main Building at University of Naples are reported and the potentialities of such algorithm as engine of smart SHM system are described. NOMENCLATURE ( ) [ ] ω j H Frequency Response Function matrix { }r ψ Mode shape vector for the r-th mode r λ System pole corresponding to the r-th mode ω Circular frequency r Q Scaling factor for the r-th mode j Immaginary unit ( ) 1 − = j INTRODUCTION In recent years a growing interest in systems and techniques for early detection of damage based on vibration analysis has raised. Vibration-based techniques aims at tracking changes in structural response directly or indirectly related to the mechanical characteristics of the structure before and after damage, i.e. natural frequencies and mode shapes. However, changes of environmental and operational conditions make structural health monitoring complex and affect modal parameters as well [1, 2]. An alternative approach is based on the post-processing of measures to detect anomalies from recorded time histories (ARMAV modelling, wavelet decomposition, etc.). In both cases, a common trend is to develop methods able to automate detection process and to exploit recent advances in information technologies (IT) [3]. A relevant aspect related to the applicability of damage detection techniques as a part of monitoring practices is an automated identification and tracking procedure. This is not a trivial task since traditional modal identification always requires extensive interaction from an experienced user. Nevertheless, computational loads have to be taken into account in order to evaluate the applicability of modal identification techniques for damage detection purposes. In fact, fast on-line data processing is crucial for quickly varying in time systems (such as a rocket burning fuel). However, a number of vibration-based condition monitoring applications are performed at very different time scales resulting in satisfactory time steps for on-line data analysis. Proceedings of the IMAC-XXVII February 9-12, 2009 Orlando, Florida USA ©2009 Society for Experimental Mechanics Inc. Interesting examples are related to structural monitoring of large structures such as bridges [2, 4] or offshore platforms [4, 5]. In the present paper, a procedure, implemented in LabView environment and able to overcome some typical drawbacks of classical operational modal analysis, is described. The algorithm and the software are briefly discussed and a specific case study, related to the integration of the software into fully automated SHM systems, is analyzed. REVIEW OF AUTOMATED MODAL IDENTIFICATION ALGORITHMS The last few years have seen a large effort in the development of vibration based damage detection techniques [4]. In fact, since the dynamic behaviour of a structure is influenced by damage, it is possible to detect occurrence of relevant damage levels through the evolution of modal parameters [6, 7]. However, changes in environmental and operational conditions can affect the modal parameters estimation [8] as well. In this framework, an automated identification and tracking procedure is a fundamental step, because traditional modal identification requires extensive interaction from an experienced user [9]. Currently, there are some advancements in this field, with the development of methods based on control theory (both in time and frequency domain) and methods based on conventional signal processing. As methods based on control theory are concerned, the model order is usually over-specified to get all physical modes present in the frequency range of interest according to classical modal analysis. However, physical and mathematical poles have to be distinguished. This practice requires large interaction with an expert user [10] and effective tools like the stabilization diagram. Selection of physical poles is not a trivial task: it may be difficult and time-consuming depending on the quality of data, the performance of the estimator (even if there are interesting advancements in this field [11]) and the experience of the user. Extensive interaction between tools and user is basically inappropriate for monitoring purposes. The first proposal for automated modal identification was based on the Least Square Complex Frequency (LSCF) method [9]. In this case the selection of physical poles from a high order model is based on a number of deterministic and stochastic criteria and a fuzzy clustering approach. However, the algorithm for pole selection is quite complex and computational demanding. In 2007 Deraemaeker et al. [12] proposed an automated operational modal analysis procedure based on the Stochastic Subspace Identification (SSI) technique. It is suitable as tracking method but it always requires user interaction because an initial set of modal parameters, using stochastic subspace identification and the stabilization diagram, has to be identified before launching the tracking procedure. Andersen et al. [13], instead, proposed in 2007 a fully automated method for extraction of modal parameters adopting the SSI technique. It is based on the clear stabilization diagram obtained according to a multipatch subspace approach: poles extraction is carried out by the graph theory. This algorithm seems to be very fast, so that it can be used for a monitoring routine, but further work is still needed in order to improve the numerical efficiency of the method. As the methods based on conventional signal analysis are taken into consideration, Guan et al. [14] proposed in 2005 the so-called Time Domain Filtering method, which is a tracking procedure based on the application of a band-pass filter to the system response in order to separate the single modes in the spectrum. However, the frequency limits of the filter are static and, above all, user-specified according only to the Power Spectral Density (PSD) plots of the response signals: if excitation is unknown, it is sometimes difficult to identify the regions where certain modes may be located according only to power spectrum plots. Moreover, in the case of close modes, it is very difficult, or even impossible, to correctly define such limits in a way able to follow the natural changes in modal frequencies. Finally, in 2007 Brincker et al. [15] presented an algorithm for automation of the Frequency Domain Decomposition procedure in order to remove any user interaction and use it as modal information engine in a SHM system. It is based on the identification of the modal domain around each identified peak in the singular value plot according to predefined limits for the socalled modal coherence function and modal domain function. A good initial value for such limits would be 0.8. However, if the limit value for the modal coherence indicator is somehow justified [15] depending on the standard deviation of the correlation for random vectors and of the number of measurement channels, a few indications are reported for the modal domain indicator. THE AUTOMATED MODAL IDENTIFICATION ALGORITHM In this section the main ideas underlying the procedure for automated modal parameter identification and some implementation details are reported. A more detailed description of theoretical background of the algorithm and results of its application to a number of different case studies can be found elsewhere [16, 17]. The algorithm starts from the Singular Value Decomposition (SVD) of the output Power Spectral Density (PSD) matrix. The latter is the core of the Frequency Domain Decomposition (FDD) method. By recalling that, when just a mode is dominant, in its bandwidth the Frequency Response Function (FRF) can be approximated as: ( ) [ ] { } ( ){ } T