Combined Multi-Scale Sensing and Neural Network Data Fusion for Damage Identification

Abstract

Multi-scale measurements, i.e. measurements of strain, strain gradient and integrated strain data, throughout a structural volume have demonstrated a great potential for improved damage identification. However, the large number of data and their different forms make fusion of the data difficult. To overcome this problem, a neural network data fusion approach is proposed. A simulation of damage identification in an isotropic cracked plate is presented. The crack position, angle and crack length are used as test parameters to be determined. A back-propagation neural network is trained to reproduce the crack angle and length as a function of all sensor responses. The improvement gained by using both multi-scale sensing and neural network data fusion for this specific case is significant. Testing of the sensitivity of the method to measurement errors or missing data demonstrated the robustness of the neural network to errors.