یک شبکه مبتنی بر مدل سازی های عصبی و تجزیه و تحلیل حساسیت ضریب نسبت آسیب
|کد مقاله||سال انتشار||مقاله انگلیسی||ترجمه فارسی||تعداد کلمات|
|26497||2011||9 صفحه PDF||سفارش دهید||6222 کلمه|
Publisher : Elsevier - Science Direct (الزویر - ساینس دایرکت)
Journal : Expert Systems with Applications, Volume 38, Issue 10, 15 September 2011, Pages 13405–13413
The level of structural damage after an earthquake can often be expressed using the damage ratio (DR) coefficient. This coefficient can be calculated using different formulas. A previously valorised new original formula for damage ratio derived for regular structures is implemented. This formula uses the structure response parameters of a single degree of freedom (SDOF) model. The structure response parameters of the SDOF model are obtained by analyzing a large number of non-linear numeric structure responses using earthquakes of different intensities as load input. In this paper, a multilayer perceptron (MLP) neural network is used to model the relationship between the structure parameters (natural period, elastic base shear capacity, post-elastic stiffness and damping) of an SDOF model and the damage ratio (DR) coefficient. The influence of the individual structure parameters on the damage level of a structure is then determined by performing a sensitivity analysis procedure on the trained MLP neural network.
Structural damage evaluation is an important aspect in the assessment of the inelastic response of reinforced concrete structures subjected to large alternate actions. The nature and amount of structural damage depends on the quality of the materials that compose the structural and non-structural elements, on the configuration and type of structural systems, and on the nature of the loads acting on the structures. Until recently, the damage was basically defined in qualitative terms, normally through the definition of the probable localization of such damage in a structure. The problem of damage quantification is complex, and there are not yet defined criteria for the definition of the analytical models and for the description of the damage itself. This fact is due essentially to the great variety of structural types and their properties, as well as to the characteristics of the actions. A significant development in the establishment of damage models has recently been made. These models assess the structural damage in quantitative terms (Comité Euro-International du Béton, 1996). The methods for determining the degree of damage to a building after an earthquake can be divided into empirical, subjective, and theoretical. Empirical methods are based on statistical observation of damage to buildings hit by earthquakes, while the subjective method is based on personal experience of the person who evaluates the degree of damage. Theoretical methods are based on a detailed analysis of dynamic models of structures exposed to action of one or more earthquakes. The basis for this method is to use one or more response parameters to define the level of construction defects. The outcome is a damage function, an analytical expression that defines the dependence of the degree of damage and the selected structure response parameter. Several approaches for the structural damage evaluation have been proposed in the literature as well as reports which critically review the different assessments (Coelho, 1992, Kappos, 1992, Morić, 1985 and Powell and Allahabadi, 1988). Seismic resistance analysis is basically the analysis of the damage ratio (DR) coefficient, a parameter which defines the level of structural damage. In economic terms, this coefficient actually represents the ratio of funds needed for the rehabilitation of structures damaged by earthquake and the resources necessary for the construction of an identical structure. Based on some known damage models, a valorised new original deterministic declaration of the damage ratio, DR, is given in Morić, Hadzima, and Ivanušić (2003). They propose that the seismic response analysis of regular structures (structure with symmetric plans and constant vertical stiffness) is acceptable if it is done as a simplified non-linear dynamic analysis with the time history function of ground motion as input load, and a single degree of freedom (SDOF) model with known weight, elastic stiffness, damping, elastic base shear capacity and post-elastic stiffness representing the structure. The structure response parameters of an SDOF model, obtained by analyzing a large number of non-linear numeric structure responses using earthquakes of different intensities and dominant frequencies as load input, are input into this previously valorised new original deterministic declaration of damage ratio (DR), thereby interpreting the level of structure damage at the end of the earthquake. Neural networks are used to model the relationship between the structure parameters (natural period, elastic base shear capacity, post-elastic stiffness and damping) and the damage level. This data is obtained from a databank of damage ratios, grouped by ground motion. A sensitivity analysis procedure is then applied to identify, qualitatively, the structure parameters that have a greater influence on the damage level. This would provide useful information as to the influence of the individual structure parameters on the damage level of a structure based upon an earthquake load. The rest of this paper is organized as follows: a brief description of the new original deterministic declaration of damage ratio, DR, provided by Morić et al. (2003) is given in Section 2. A classification of the structures used in the experiment is provided in Section 3. In Section 4, a brief introduction to neural networks is given as well as the structure of the neural network used. An explanation of how sensitivity analysis can be performed using neural networks is also given. Section 5 includes the results obtained after modeling the relationship and performing sensitivity analysis. A brief conclusion is provided in Section 6.
نتیجه گیری انگلیسی
Using the structure parameters of 2250 various SDOF structures and the ground acceleration time-histories of 20 real earthquakes, a data bank of 45,000 calculated values of damage ratios is created by implementing a previously valorised new original deterministic declaration of damage ratio (DR). A multilayer perceptron (MLP) neural network is successfully used to model the function relating the structure parameters (natural period, elastic base shear capacity, post-elastic stiffness and damping) of an SDOF model and the damage ratio (DR) coefficient for each earthquake. With the aid of the MLP neural network, a sensitivity analysis procedure is applied to identify, qualitatively, the input parameters that have a greater influence on the damage level. This is performed by using different input combinations of the input parameters of the network and retraining the network. For each combination of input parameters, the neural network is trained 15 times, each time with different weight initialization. The average network error (MSE) of all training sessions is taken as the network error of the neural network model. The network error is compared with the error obtained when all inputs were used. In general, using predictive importance, it is noticed that the most important parameter is the yield base shear capacity (SY) followed by natural period (T0). The post-elastic stiffness (K2) and damping of a structure (ξ) are the least important parameters. However, for catastrophic earthquakes (earthquakes having a maximum peak acceleration greater than 0.6g), the most important parameter is natural period (T0) followed by yield base shear capacity (SY). The post-elastic stiffness (K2) and damping of a structure (ξ) are the least important parameters.