Damage Identification Using Measured and Simulated Guided Wave Damage Interaction Coefficients Predicted Ad Hoc by...

Thin-walled structures are widely used in aeronautical and aerospace engineering due to their light weight and high structural performance. Ensuring their integrity is crucial for safety and reliability, which is why structural health monitoring (SHM) methods, such as guided wave-based techniques, have been developed to detect and characterize damage in such components. This study presents a novel damage identification procedure for guided wave-based SHM using deep neural networks (DNNs) trained with experimental data. This technique employs the so-called wave damage interaction coefficients (WDICs) as highly sensitive damage features that describe the unique scattering pattern around possible damage. The DNNs learn intricate relationships between damage characteristics, e.g., size or orientation, and corresponding WDIC patterns from only a limited number of damage cases. An experimental training data set is used, where the WDICs of a selected damage type are extracted from measurements using a scanning laser Doppler vibrometer. Surface-bonded artificial damages are selected herein for demonstration purposes. It is demonstrated that smart DNN interpolations can replicate WDIC patterns even when trained on noisy measurement data, and their generalization capabilities allow for precise predictions for damages with arbitrary properties within the range of trained damage characteristics. These WDIC predictions are readily available, i.e., ad hoc, and can be compared to measurement data from an unknown damage for damage characterization. Furthermore, the fully trained DNN allows for predicting WDICs specifically for the sensing angles requested during inspection. Additionally, an anglewise principal component analysis is proposed to efficiently reduce the feature dimensionality on average by more than 90% while accounting for the angular dependencies of the WDICs. The proposed damage identification methodology is investigated under challenging conditions using experimental data from only three sensors of a damage case not contained in the training data sets. Detailed statistical analyses indicate excellent performance and high recognition accuracy for this experimental data-based approach. This study also analyzes differences between simulated and experimental WDIC patterns. Therefore, an existing DNN trained on simulated data is also employed. The differences between the simulations and experiments affect the identification performance, and the resulting limitations of the simulation-based approach are clearly explained. This highlights the potential of the proposed experimental data-based DNN methodology for practical applications of guided wave-based SHM.