Direct identification of structural parameters from dynamic responses with neural networks

A novel neural network-based strategy is proposed and developed for the direct identification of structural parameters (stiffness and damping coefficients) from the time-domain dynamic responses of an object structure without any eigenvalue analysis and extraction and optimization process that is required in many identification algorithms for inverse problems. Two back-propagation neural networks are constructed to facilitate the process of parameter identifications. The first one, called emulator neural network, is to model the behavior of a reference structure that has the same overall dimension and topology as the object structure to be identified. After having been properly trained with the dynamic responses of the reference structure under a given dynamic excitation, the emulator neural network can be used as a nonparametric model of the reference structure to forecast its dynamic response with sufficient accuracy. However, when the parameters of the reference structure are modified to form a so-called associated structure, the dynamic responses forecast by the network will differ from the simulated responses of the associated structure. Their difference can be assessed with a proposed root mean square (RMS) difference vector for both velocity and displacement responses. With the associated structural parameters and their corresponding RMS difference vectors, another network, called parametric evaluation neural network, can be trained. In this study, several 5-story frames are considered as example object structures with simulated displacement and velocity time histories that mimic the measured dynamic responses in practice. The performance of the proposed strategy has been demonstrated quite satisfactorily; the error for the estimation of each stiffness or damping coefficient is less than 10% even in the presence of 7% noise. Numerical simulations show that the accuracy of the identified parameters can be significantly improved by injecting noise in the training patterns for the parametric evaluation neural network. The proposed strategy is extremely efficient in computation and thus has potential of becoming a practical tool for near real time monitoring of civil infrastructures.