Robustness of wavelet features for damage detection with neural network and digital twin

Recent studies demonstrate that the wavelet energy–based features are highly sensitive to local damages, and accurate damage identification could be achieved by integrating such features with neural networks. However, the robustness of the features in practical applications and the feasibility of generating damage information with a digital twin to facilitate training of neural networks have not been systematically investigated. This paper aims to provide new insight into the practicality of using wavelet energy–based features for accurate damage detection of structures. A systematic investigation into the tolerance of the normalised wavelet packet node energy features with neural network (NWPNE-NN) approach against measurement noises, excitations uncertainties and limited frequency range in the measured responses is carried out, with consideration of data fusion from multiple measurement points. The feasibility of a digital twin to simulate damage scenarios and generate training data for neural networks is explored, and the use of relative WPNE as a new feature is proposed. The results show that the NWPNE-NN approach is capable of detecting and quantifying the main damage in a structure. The neural networks trained by data generated from a well-calibrated digital twin is capable of identifying damage in the actual structure.