Machine learning inverse surrogates for damage identification in plates based on Lamb waves

This work proposes a black-box purely data-driven strategy for structural damage identification concerning localized damaged regions in plate-like structures. The proposed strategy is based on Convolution Neural Networks in the framework of a supervised learning regression task. The relationships between the estimated and target damage parameters are investigated, focusing on the physical interpretability of the damage recovery results. The positional parameters are found to be much more easily estimated than those describing damage size and damage intensity, in accordance with the literature where similar parameterizations are considered. The high accuracy with which it is possible to estimate the positional parameters explains, in part, the many successful approaches found in literature where damage localization is treated as a classification problem. Some numerical analyses are shown for a convolutional neural network architecture with diverse damage scenarios in an elastic plate considering Lamb waves, three actuators and 16 sensors. The first surrogate inverse model training considers homogeneous material properties and the second one considers non-homogeneous material properties. Results are evaluated using a set of overlap metrics, which help identify both the accuracy and the limitations of the inverse surrogates in damage recovery. The inverse surrogate trained with non-homogeneous material properties proved robust with respect to system variability.