Article Preview
TopIntroduction
The dynamic response of embankments and dams was largely studied by using one-dimensional, two-dimensional or three-dimensional models (Yu et al., 2005; Singh et al., 2005; Siyahi & Arslan, 2008, Trana et al., 2009; Sanchez Lizarrag & Lai, 2014). Whereas discrete models such as finite elements and sometimes combined with boundary elements receive more interest by researchers, simple models continue to be used considering their simplicity and effectiveness in much of treated cases (Papadimitriou et al., 2014; Andrianopoulos et al., 2014; Neamaty & Khalili, 2015). In other way, reduction of seismic risk in areas with strong seismicity needs microzonation studies which must include mechanical parameters of soil stratums. These parameters are classically obtained from laboratory and/or in situ tests. However, these classical techniques of investigation (drilling and sampling), in situ tests, or geophysical means are expensive and time consuming (Güllü, 2013).
For this purpose, an inverse analysis is proposed to identify mechanical parameters of soil stratums and earthen dams by using three kinds of optimization algorithms to minimize the error function between measured and predicted spectral values in conjunction of accelerometer data recorded at free surfaces of soils and rocks. The first one is the local traditional Levenberg - Marquart gradient-based search method, the second one is based on evolutionary genetic algorithms which belong to the global methods, and the third one is hybridization between the genetic algorithm scheme at the beginning of the optimization process, in order to improve the initial guessing of the parameters, and the Levenberg-Marquardt method with the aim to accelerate the convergence in the final phase of the identification procedure.
Indeed, system identification and inverse problem techniques have been used in estimation of soil parameters, using experimental and observational earthquake data (Zeghal & Abdel-Ghaffar, 2009; Oskay & Zeghal, 2011; Foti & Parolai, 2011; Kunimatsu et al., 2011; Yamanaka et al., 2016). Traditionally, system identification is performed by using analytic models, i.e. obtaining mathematically the transfer function (amplification function) or using experimental input–output data (Harichane et al., 2005). In the last way, the identification can be achieved in non-parametric (Ching & Glaser, 2003) or parametric ways. Most of the parametric techniques used in the literature to solve optimization problems are gradient-based search methods. These optimization search techniques have been successfully applied to find the dynamic parameters related with the one-dimensional (1D) wave propagation theory (Martin et al., 2010; Pecha et al., 2012; Harichane et al., 2012). However, these optimization procedures need a previous knowledge about the parameters range for constructing the error function to optimize.