Selection of input vectors to neural networks for structural damage identification

This paper addresses constructing appropriate input vectors (input patterns) to neural networks for hierarchical identification of damage location and extent from measured modal properties. Hierarchical use of neural networks is feasible for damage detection of large-scale civil structures such as cable-supported bridges and tall buildings. The neural network is first trained using one-level damage samples to locate the position of damage. After the damage location is determined, the network is re-trained by an incremental weight update method using additional samples corresponding to different damage degrees but only at the identified location. The re-trained network offers an accurate evaluation of the damage extent. The input vectors selected for this purpose fulfil the conditions: (a) most parameters of the input vectors are arguably independent of damage extent and only depend on damage location; (b) all parameters of the input vectors can be computed from several natural frequencies and a few incomplete modal vectors. The damage detection capacity of such constructed networks is experimentally verified on a steel frame with extent-unknown damage inflicted at its connections.