Reconstruction of 2D Guided Wave Propagation Data Using Generative Adversarial Networks

Abstract

Guided wave propagation analysis is one of the critical areas in Structural Health Monitoring (SHM) that enables the detection and assessment of damage in critical infrastructure. Visualization of wave interactions with structural anomalies, such as holes or cracks, gives important insights into the integrity and behavior of materials. In particular, automatic decision systems taking guided wave data as input can be used to monitor vital infrastructure, provided that they are trained on large amounts of diverse data. Unfortunately, due to the cost, complexity, and time-consuming nature of experimental measurements, obtaining high-quality labeled data for guided wave propagation analysis remains a significant challenge. These restrictions on generating diverse datasets often make it more difficult to train reliable machine learning models for SHM, especially when handling a variety of complex damage scenarios. Generative Adversarial Network (GAN) can offer a solution to those issues by producing realistic synthetic datasets that replicates key features of experimentally obtain data, allowing for cost-effective augmentation of SHM datasets. Those generative models are well suited for this task, based on their ability to simulate accurate representations of guided wave propagation through damaged structures, capturing subtle interactions between waves and structural anomalies. This allows for augmentation of SHM datasets at a reasonable cost, which eventually improves machine learning models ability to generalize damage scenarios that haven’t been seen and reducing dependency on expensive experimental setups. In this paper we demonstrate how to recreate two-dimensional (2D) guided wave propagation data through the aluminum plate with damages modeled as flat-bottom holes. Our method produces high-fidelity images with very close resemblance to training data obtained using laser vibrometer. This approach highlights the potential of generative models to improve data diversity and availability for SHM applications, in particular in simulation of previously unseen damage cases which enable directed augmentation of existing SHM databases. This strategy can reduce the gap between limited experimental datasets and machine learning-driven SHM systems.