Hierarchical development of training database for artificial neural network-based damage identification

Abstract

Though serving as an effective means for damage identification, the capability of an artificial neural network (ANN) for quantitative prediction is substantially dependent on the amount of training data. In virtue of a concept of “ Digital Damage Fingerprints” (DDF), a hierarchical approach for the development of training databases was proposed for ANN-based damage identification. With the object of exploiting the capability of ANN to address the key questions: “ Is there damage?” and “ Where is the damage?”, the amount of training data (damage cases) was increased progressively. Mutuality was established between the quantity of training data and the accuracy of answers to the two questions of interest, and was experimentally validated by identifying the position of actual damage in carbon fibre-reinforced composite laminates. The results demonstrate that such a hierarchical approach is capable of offering prediction as to the presence and location of damage individually, with substantially reduced computational cost and effort in the development of the ANN training database.