Synthetic data generation of ultrasonic guided wave in thin plates using generative adversarial networks

Structural health monitoring using AI-assisted signal processing models has recently demonstrated high efficiency and practicality for applications that are difficult to address using traditional signal processing techniques. However, these models typically require large, richly annotated damage datasets that are time-consuming, expensive, and often impractical to obtain, making data scarcity a critical bottleneck in the field. In this work, we present a synthetic data generation framework for UGW signals in thin plates based on a conditional WGAN with Gradient Penalty. The proposed framework generates realistic multi-channel UGW signals using only a limited number of experimental data. UGW data were collected from an aluminum plate instrumented with a sparse rectangular network of PZTs. In the first study, model was trained using only 10% of clean UGW signals, while the remaining 90% were synthetically constructed under different defect scenarios, achieved with experimental data, with correlation coefficients exceeding 0.95 and SSIM values above 0.97. In the second study, the framework successfully generated signals for previously unseen sub-zones within the same inspection region, and in the third study, the learned wave propagation characteristics were evaluated on an unseen monitoring-zone with different propagation characteristics. These results suggest that the proposed framework preserves statistically consistent spatial UGW characteristics, including temporal, and dispersive spectral energy behavior within the experimental configuration. Overall, the proposed approach substantially reduces the experimental effort while maintaining strong temporal and time-frequency consistency with experimentally measured UGWs and offering a promising solution to data scarcity in SHM.