Simulation of damage-induced wave propagation in composites using knowledge-embedded generative adversarial networks for structural health monitoring

Recent advancements in data-driven methods have greatly improved the efficiency of structural health monitoring in detecting damage and deriving knowledge from monitoring data. These methods, though, are often limited in their applicability and implementation due to the challenge of sparse monitoring data. The use of simulated data by traditional simulation approaches is an alternative. However, the accuracy of these methods is frequently constrained by oversimplification and assumptions, leading to simulations that do not match real measurements. This limits their practical usefulness and cost-effectiveness, underscoring the need for more innovative approaches to data generation in the field of structural health monitoring. This paper proposes a novel method for simulating measurement data of monitored structures, even when the structure to be simulated lacks training examples from the target domain (damaged state), by leveraging measurement data exclusively from the intact structure. The proposed method employs a knowledge-embedded generative adversarial network (KE-GAN) that incorporates domain knowledge into the generator loss to address the lack of data from the target domain. Initially, the generator is trained on data from a source domain, generally measurement data from the undamaged state, to learn the properties of the data that should be simulated. In a second phase, the loss function is enhanced with the condition that characterizes the target domain, typically specific damage to the structure. This integration enables the generation of data that closely resembles the training examples from the source domain while also satisfying the desired condition of the target domain. To evaluate the proposed method, a case study was conducted to generate high-frequency guided wave propagation in a carbon fiber-reinforced polymer plate. Twelve piezoelectric transducers were placed along the plate measuring the wave amplitude at the corresponding placement. The KE-GAN was trained exclusively on intact state data, and the Pearson correlation coefficient between baseline and generated data were used as an indicator of damage location. The generated data were then compared to the measured data of the damaged state by calculating the mean square error and through comparison of the output of the reconstruction algorithm for probabilistic inspection of damage. The results demonstrate the KE-GAN’s ability to accurately simulate various damage scenarios with minimal deviation from actual measured data, even with limited availability of baseline training data. This underscores its potential for generating measurement data in domains where comprehensive datasets are scarce.