Bayesian regularized back propagation neural network model for seismic damage prediction on multistory masonry buildings

In this paper, Bayesian Regularized Back Propagation Neural Network (BRBPNN) technique was applied in seismic damage prediction on buildings. The seismic damage influences factors are summarized as nine input parameters for the network. They are site condition X-1, integrity of building X-2, building shape X-3, building height X-4, mortar grade of wall at bottom X-5, area ratio of brick wall X-6, X-7, building format X-8 and seismic intensity X-9. From earthquakes of Langeang-gengma (1988), Wuding (1995), Lijiang (1996) and Tanshang (1978), 50 seismic damage samples for multistory masonry buildings were collected. After checking the square weights of the input layers of the BRBPNN models, the factor X-3 is found to be not efficient. By contrasting the Mean Square Errors (MSE) on the training and testing samples of different models, the model 8-5-5 is recommended. Increase of the neurons in the hidden layer has little influence to the efficiency of the neural network. The importance of the factors is listed as X-9, X-7, X-6, X-4, X-8, X-2, X-1, and X-5. The scatter diagrams show that the effect of each factor on the seismic damage of multistory masonry buildings. The factor like X-9, X-1, X-2, and X-8 are more dependent for the problem, although they all seem to be nonlinear with the seismic damage. Results show that Bayesian Regularized Back Propagation Neural Network is of automated regularization parameter selection capability and thus may ensure the excellent generation ability and robustness. This study threw some light on the applicability of seismic damage prediction on the multistory masonry buildings.