This study presents an application of Artificial Neural Network (ANN) technique for the damage detection of El-Ferdan Bridge, Egypt. A finite element model of the bridge was constructed using the actual member sizes obtained from as-built drawings. Damage in a specific member was artificially simulated by reduction of the member cross-sectional area. Several damage conditions were assumed to cause reduction of 20, 40, 60, and 80 percent of the member's area and to occur in different members of the bridge. In each case, the corresponding static and dynamic properties of the bridge were obtained and fed to generalized feed forward ANNs with different learning rules. Several networks were designed to have different sets of network inputs including the static deflections of the bridge dick; the dynamic characteristics in terms of modal time periods of vibration; and a combination of both characteristics. In all cases the output layer consisted of two nodes representing the location of damage (member name), and the percentage of damage in that member. In each of these networks, several network topologies were examined including networks with one hidden layer and networks with two hidden layers, containing 5, 10, 15, 20, 25, 30, 35, 40, 45, and 50 neurons per layer. Sensitivity analysis results indicated the importance of the dynamic characteristics of the bridge for successful ANN predictions. Most successful network resulted in a 90.74% success in the prediction of the damage location together with a mean square error of 0.043 for the prediction of damage percentage. It is concluded that the ANN technique is very economic and accurate to be considered as an extra alternative for damage detection or structural health monitoring of large structures. |
