LAUSR

Abstract Structural damage assessment is a challenging problem of study due to lack of information in data measurement and the difficulty of extracting noisy features from the structural responses. Therefore, this paper proposes an effective deep feedforward neural networks (DFNN) method for damage identification of truss structures based on noisy incomplete modal data. In the proposed approach, incomplete datasets are randomly generated by a reducing finite element (FE) model. Based on the collected data, the DFNN model is constructed to predict damage position and severity of structures. To obtain a better performance of the network, the new ReLu activation function and Adadelta algorithm are employed in this work. In addition, the state-of-the-art mini-batch and dropout techniques are adopted to speed up the training process and avoid the over-fitting issue in training networks. Various hyperparameters such as number of hidden units, layers and epoches are surveyed to built a good training model. In order to demonstrate the efficiency and stability of the proposed method, a 31-bar planar truss structure and a 52-bar dome-like space truss structure are investigated with various damage scenarios. Moreover, the performance of the DFNN method is not only illustrated with the noise free input data but also with noisy input data. Different noise levels of the input data are taken into account in this study. To accurately predict the damage location and severity of the structures, 10000 and 20000 data samples corresponding to the 31-bar planar truss and the 52-bar dome-like space truss are randomly created in term of quantity of damage members, damage locations and damage severity of the structures for training the DFNN models. The results predicted by the DFNN using incomplete modal data are compared with those of the complete and actual models. The obtained results indicate that the DFNN is a promising method in damage localization and quantification of civil engineering structures.