A novel wavelet energy feature for damage identification with a digital twin considering measurement uncertainties

In structural health monitoring and damage identification literature, supervised machine learning has been commonly adopted. However, the establishment of the training dataset remains to be an open question. Besides indirect experimental methods such as adding masses, the use of a digital replica (digital twin) in a reference, undamaged state is deemed to be a necessity, so that a variety of future damaged states may be generated by varying the properties of the digital twin. However, little research has been available in the literature that addresses the challenge of the modelling errors in such an approach. This study advances the digital-twin-based damage identification approach by examining the ability of a digital twin to generate the wavelet packet node energy (WPNE) features for a variety of damage states and identifying the influences of inherent uncertain physical properties, particularly damping. A novel WPNE feature is developed through feature engineering, effectively mitigating inaccuracies brought about in the damping estimates. The proposed digital-twin-based damage identification approach with the new WPNE feature is validated via numerical and laboratory experiments, demonstrating its robustness against inevitable modelling errors. This work brings the role of digital twins in damage identification a step further towards real-life applications.