Damage Detection in Structures Using Artificial Neural Networks

In order to select a sensitive input parameter for artificial neural networks in damage detection and construct an efficient and robust back propagation algorithm for damage assessment, the application of neural networks to damage detection in structures is summarized and analyzed in this paper. By discussing the use of natural frequency as a diagnostic parameter, natural frequency can rationally reflect damage location but not provide enough information about damage degree. Mode shape and transfer function include abundance information about damage degree compared with natural frequency but have a large measurement error. And three improved back propagation algorithms that are adaptive variable step-size algorithm, Levenberg-Marquart algorithm and homogeneous algorithm are introduced. The result shows that Levenberg-Marquart algorithm harmonizes Gauss-Newton method with steepest descent method and tunes gradually to Gauss-Newton method when the result can not converge to the minimum. Thus choosing complete vibration modal parameters and using Levenberg-Marquart algorithm, structural damage can be effectively detected.