چكيده به لاتين
Structures during their lifetime may experience various loads, including earthquakes, which will damage the structures. For this reason, identifying and evaluating damage to structures has been one of the issues of concern to engineers to prevent a lot of financial and human losses. Therefore, engineers have always been looking for the best and most effective ways to identify damage to structures. In recent years, the ability of neural networks in various fields such as pattern recognition and feature extraction from data has been considered by engineers to solve the problems of damage detection and structural health monitoring. This thesis examines the ability of artificial neural networks and deep learning to detect damages in reinforced concrete structures. For this purpose, first, by using a feedforward neural network, which is a static network, damage in reinforced concrete structures has been identified. In this method, using specific data obtained from acceleration, the input parameters required by the network are provided, and using deductive concepts of the relationship between displacement and residual drift with structural damage, the damage index is defined and structural damage is identified. Then, in another method, using the long short-term memory (LSTM) network, which is a dynamic network and can consider the long-term dependencies of the time series data, the center of mass acceleration data are given to the network and re-examined damage identification in the structure as before. For this purpose, reinforced concrete structures with regular and irregular geometric plans with different heights have been designed and then subjected to nonlinear time history analyses. Then, the data required for training networks are extracted, and appropriate pre-processing of each network is done on the data to achieve good results with the least computational complexity. Examination of the results also showed that both networks can detect damage with an accuracy of over 85%. On the other hand, it was observed that the long short-term memory network has higher accuracy and better performance in predicting damage levels when acceleration is applied to the structures in both directions. Also, if the damage levels are reduced from 5 to 3 levels, the accuracy of both networks will reach over 90%.
