Constructing input vectors to neural networks for structural damage identification

Abstract

This paper addresses construction of appropriate input vectors (input patterns) toneural networks for hierarchical identification of structural damage locationand extent from measured modal properties. Hierarchical use of neuralnetworks is feasible for damage detection of large-scale civil structures such ascable-supported bridges and tall buildings. The neural network is first trainedusing one-level damage samples to locate the position of damage. After thedamage location is determined, the network is re-trained by an incremental weightupdate method using additional samples corresponding to different damagedegrees but only at the identified location. The re-trained network offers anaccurate evaluation of the damage extent. The input vectors have been designedto meet two requirements: (i) most parameters of the input vectors are arguablyindependent of damage extent and only depend on damage location; (ii) allparameters of the input vectors can be computed from several natural frequenciesand a few incomplete modal vectors. The damage detection capacity of suchconstructed networks is experimentally verified on a steel frame withextent-unknown damage inflicted at its connections, and the applicability of thehierarchical identification strategy to cable-supported bridges is discussed.