Download citation:
BibTeX | RIS | EndNote | Medlars | ProCite | Reference Manager | RefWorks
Send citation to:

---

These days the accurate estimation of seismic demand and capacity of structures are truly significant in the field of performance based earthquake engineering. Several methods exist to determine these parameters such as non-linear time history analysis and Incremental dynamic analysis (IDA). Because the history of seismic accelerogram records refers to the current century, in some areas there still exists no appropriate seismic record to perform the analyses; therefore in these cases we need to generate artificial accelerograms. In this paper a new combinational method is introduced to generate far-field artificial accelerograms using artificial neural network and wavelet packet transform (WPT) methods. In this method according to the geoseismic characteristics of the site and non-linear characteristics of the equivalent single degree of freedom (SDOF) system, several artificial accelerograms are generated. In order to consider the non-linear parameters to generate the accelerograms, IDA method is used. The values of intensity measure (IM) for all IDA curves are determined at specific levels of damage measure (DM) and are considered as the input data of the multilayer feed forward (MLFF) neural network. Accelerograms which are selected according to the geoseismic characteristics of the site are changed to standard forms and then decomposed using wavelet packet transform. The effective wavelet packet coefficients are selected according to an appropriate desired effective variance ratio of wavelet packet coefficient. Then, effective coefficient of each packet is considered as the output of a neural network. In order to enhance the efficiency of the network, principal components analysis (PCA) is used to reduce the number of the input data dimensions. In this paper neural network is trained by backpropagation algorithm as repetitive. After training the MLFF neural network, we should test the network for accelerograms not included in the training set. For this purpose we should use the IDA curve of each accelerogram out of the training set as the input of the neural network to generate the effective WPT coefficients. When a neural network is trained properly, we can now generate artificial accelerograms using a 50% fractile IDA curve as the input of the neural network. Adding a Gaussian random number to the output of each neuron in the neural network layers, we are able to generate several accelerograms according to 50% fractile IDA curve. In order to improve the condition of generated accelerograms according to 50% fractile IDA curve, a correction factor is used repeatedly for detail coefficients of discrete wavelet transform in jth level of generated accelerogram. Finally a SDOF system with perfectly elasto-plastic initial loading curve is used to show the efficiency of the proposed method to generate artificial accelerogram. The accuracy of this method depends on the accuracy of the trained neural networks. If the neural networks are trained appropriately with IDA curve, the generated accelerogram can estimate the IDA parameters of the SDOF system more properly. Also it is shown that suggested method can generate artificial accelerograms with frequency content almost close to the initial earthquake records.

Keywords: Artificial Accelerogram Generation, Wavelet Packet Transform, artificial neural networks, Incremental dynamic analysis

Full-Text [PDF 501 kb]   (7798 Downloads)    

Add your comments about this article : Your username or Email:

---