Effect of Input Variable for Neural Network Architecture in Predicting Building Damage Subjected to Earthquake

Abstract

The artificial intelligent methodology is applied to building inspection system in this paper. Inspection results from Applied Technology Council (ATC) procedures are used as an indicator to the building damage level. Backpropagation algorithm with one hidden layer is used to develop the neural network and Borland C++ is used as the programming language. The choice of input variables is a fundamental. The neural network performance is analysed by removed the input one by one from the network. The mean square error (MSE) value is equal to 0.027 when all inputs are applied to the network. The lowest MSE is 0.025 when the length is excluded from the network. Lower MSE value indicates lower error among others. Even though the lowest MSE is presented when the length is excluded from the network, but the different was too small which is 0.002. Thus, the length is still applied in network. From the application of Artificial Neural Network (ANN), building samples gave the lowest value of MSE equal to 0.027 when 15 hidden neurons applied and the highest linear correlation coefficient was obtained when r is equal to 0.839 in testing phase and 0.762 in validation phase. Out of 112 samples in the testing phase, 104 samples were predicted accurately for degree of damage rating for the building which represents 93 % from the total data used. In the validation phase, 39 out of 52 samples were predicted accurately with 75 % accuracy. The results of this study indicate that the ANNs provide an efficient means of damage forecasting and would be useful by the owners of the building to predict building conditions under seismic load.