Skip to Main Content
A neural network (NN)-based technique making direct use of measured dynamic responses in civil structures is proposed to model the structure and detect eventual anomalies with their location and extent. Although numerous researches were conducted to apply NN for damage detection purposes, the problem constituted by the selection of an appropriate architecture for the networks still remains a major obstacle impeding their applicability. In order to avoid this shortcoming, the proposed algorithm performs the modeling of the structure stepwise by successive integration-like neural operations, which permits to reduce effectively the size of the networks and simplify effectively their architecture. The damage parameter is decided to be the restoring forces and corresponding stiffness of each major structural member. The trained network fed with data of the structure encountering diverse damage events under various loading episodes reconstructs the actual restoring force loops and the ones that should be obtained for the undamaged structure, of which comparison provides accurate estimation of damages. A shear building example verifies the efficiency and accuracy of the proposed method in detecting, locating and giving the extent of damages in real time.