Data-driven damage identification with wavelet energy features

Abstract

Vibration-based damage identification is an important topic in structural health monitoring (SHM). Some recent studies have demonstrated that the wavelet energy features contain detailed information which can be highly sensitive to local damages, and in conjunction with supervised machine learning, promising results have been exhibited in terms of both sensitivity and accuracy. However, the susceptibility of wavelet energy features to uncertainties in real measurement conditions has not been systematically investigated. The crucial question as how the training data would be obtained in supervised learning environment has not been properly addressed.

Furthermore, the effectiveness of using wavelet energy features for identification of multiple damages, and how this should be implemented, has not been explored.

This thesis aims to address the abovementioned challenges and advance wavelet energy features - machine learning approach towards real-life application. At the first stage, the tolerance of machine learning models using wavelet packet node energy (WPNE) as damage sensitive features to measurement uncertainties is investigated, particularly in terms of variations in the input excitation and measurement noises, using both numerical simulation and experimental studies. In addition, a dataset obtained from testing of a steel pedestrian bridge is used to further validate the viability of the approach in real-life measurement conditions. The effectiveness of data fusion of multiple sources in neutralising the interferences is discussed. A digital twin aided wavelet neural network hybrid approach is then explored. In the establishment of the digital twin, fusion of data from multiple sources is particularly investigated.

Subsequently, the effect of modelling errors, particularly in damping, on the performance of WPNE in damage identification is investigated. A novel damage sensitive feature based on the linear relationships between particular WPNEs, called RSWPNE, is proposed. The RSWPNE is demonstrated to be immune to modelling errors in damping. Numerical and laboratory experimental studies are then conducted to demonstrate the robustness of the modified WPNE features in withstanding inaccurate estimates of damping while maintaining the high sensitivity to structural changes.

For the multiple damage scenarios, a two-stage framework using ensemble binary classifiers and the novel RSWPNE is proposed. The performance of the proposed approach in the identification of arbitrary combinations of multiple damages is examined on laboratory structures. Results demonstrate that the proposed two-stage framework is effective in identifying the occurrence and severity of multiple damages.