APPROACH TO PREDICTION OF R/C BUILDINGS’ SEISMIC DAMAGE AS PATTERN RECOGNITION PROBLEM USING ARTIFICIAL NEURAL NETWORKS

Abstract eng:
In the present paper the investigation of the problem of reinforced concrete (r/c) buildings’ seismic damage prediction is exhibited utilizing Artificial Neural Networks (ANN). More specifically, the problem is formulated and solved in terms of the Pattern Recognition Problem using Multilayer Feedforward Perceptron Networks (MFP). The networks are trained by the implementation of two training algorithms: the Scaled Conjugate Gradient algorithm (SCG) and the Resilient Back-Propagation algorithm (RP). The training data-set is created by means of Nonlinear Time History Analyses (NTHA) of 30 r/c buildings which are subjected to 65 earthquakes. The selected buildings have different heights, structural systems and structural eccentricities, and are designed on the basis of the suggestions of Eurocodes. The damage index which is used to describe the seismic damage state of buildings is the Maximum Interstorey Drift Ratio (MIDR). In the context of the present paper the influence of the number and the combination of seismic parameters which describe the level of impact of seismic excitations on the r/c buildings is also investigated. To this end, 8 different and widely used seismic parameters are utilized. Furthermore, the influence of the number of hidden layers, the number of neurons in the hidden layers, as well as the activation functions of neurons is also examined. The generalization abilities of the optimum configured ANNs are investigated through the assessment of their performance in the case of prediction of seismic damage state of the selected buildings subjected to 16 earthquakes different from the earthquakes which are used in the creation of the training data-set. The most significant conclusion that turned out is that the ANNs can reliably classify the r/c buildings into pre-defined damage classes if they are appropriately configured. More specifically, the parametric investigations prove that the most important factor for the effective ANNs’ configuration is the activation functions of output layer’s neurons (tansig function) as well as the number of the hidden layers (the utilization of two hidden layers leads to better results). As regards the examination of the optimum number and combination of seismic input parameters (using the “Stepwise” sensitivity analysis method) it is proved that are dependent on the utilized training algorithm.

• Export as: BibTeX | MARC | MARCXML | DC | EndNote | NLM | RefWorks
• Add to your basket