ANN-based damage detection with two-stage approach and model reduction

Abstract

The implementation of artificial neural network (ANN) is limited in model-based damage detection of large-scale structures. A large potential damaged area impairs the convergence performance of ANN training, and numerous structural DOFs decrease computational efficiency. This article proposes a novel strategy of ANN-based damage detection with two-stage approach and model reduction. To reduce computation DOFs and narrow down the detected area, the original finite element model is divided into manageable substructures and transferred into a super-element model through substructural reduction and integration. Then, damage localization is first performed to search for suspicious damaged substructure by the constructed ANN, which is trained by the defined modal group strain energy analytically calculated from the reduced model. Then, element-stage damage detection is conducted by minimizing the defined reconstructed modal group, and more detailed damage information is desired. The proposed methods could be performed conveniently as traditional modal-based damage detection methods, while it further overcome the difficulty of deal with closely spaced modes. Furthermore, the incorporated ANN provides better timeliness for health monitoring. Numerical and experimental cases are performed for verification. Analysis for the neural network and adjustment are studied.