NEURAL NETWORKS FOR THE PREDICTION OF SEISMIC DAMAGE IN REINFORCED CONCRETE STRUCTURES

The possible seismic damage to structures can be evaluated using either vulnerability curves or non-linear dynamic analyses. However, the nonlinear dynamic analyses are time-consuming and the vulnerability curves can only be applied to a certain type of structures. There have been approaches to find correlation between several ground motion parameters and damage indices for buildings in the past (Lönhoff, 2017; Kwon, 2006; de Lautour, 2009). For this purpose, this paper deals with the idea of using artificial neural networks to significantly simplify the calculations. In fact, neural networks are able to find complex correlations in large data sets. Since the aim of theinvestigation is to determine the damage caused by an earthquake, suitable acceleration time curves must first be selected as training data. The correct selection of input data is crucial for the successful training of an artificial neural network. For the analysis of the damage, only strong earthquakes with a distance of more than 15 km were used, because near-field earthquakes show special damage behaviour due to the peak loads. In order to investigate the relationship between engineering seismological parameters and structural damage, nonlinear dynamic analyses of different structures are performed with the finite element program OpenSees. Four buildings of different heights are used to cover a wide range of common structures. The models are all made of reinforced concrete components and consist of a two-storey, a four-storey, a six-storey and an eight-storey building. To simplify matters, the buildings are analysed as 2D models. To investigate the damage, the Overall Structural Damage Index for the widely useddamage indicator Park-Ang with Equation 1 and Kunnath were calculated for the given structures for each earthquake (Park Ang, 1984). Additionally, the maximum displacement of the floors among each other (MIDR) and the maximum roof displacement (MRDR) have been calculated as global damage indicators. There have been several damage indicators investigated to verify the independence of the surrogate model from the utilized damage index.

Keywords:

Artificial neural networks , Damage indices , Earthquake , Intensity measures