Ultrasound Reconstruction Tomography Using Neural Networks Trained with Simulated Data: A Case of Theoretical Gradient Damage in Concrete
[EN] Gradient damage in cementitious materials is typically driven by chemical and/or physical agents that penetrate from the exterior. Depending on the mechanism, it may lead to reduced porosity, cracking, or steel corrosion (e.g., carbonation), or to increased porosity, microcracking, expansion, a...
| Autores: | , , , , , |
|---|---|
| Tipo de recurso: | artículo |
| Fecha de publicación: | 2025 |
| País: | España |
| Institución: | Universitat Politècnica de València (UPV) |
| Repositorio: | RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia |
| Idioma: | inglés |
| OAI Identifier: | oai:riunet.upv.es:10251/229988 |
| Acceso en línea: | https://riunet.upv.es/handle/10251/229988 |
| Access Level: | acceso abierto |
| Palabra clave: | Ultrasound tomography Reconstruction algorithm Neural network NDT Concrete damage |
| Sumario: | [EN] Gradient damage in cementitious materials is typically driven by chemical and/or physical agents that penetrate from the exterior. Depending on the mechanism, it may lead to reduced porosity, cracking, or steel corrosion (e.g., carbonation), or to increased porosity, microcracking, expansion, and spalling (e.g., external sulphate or acid attack, as well as thermal damage). Delineating the boundary of this damage is therefore essential for concrete quality assessment. This work pursues two goals. First, we use neural networks (NNs) to perform ultrasonic tomographic reconstruction of concrete specimens in order to estimate the advancing damage front. Unlike X-ray tomography, ultrasonic tomography is strongly affected by diffraction and related wave phenomena; NNs can learn to compensate for these effects but require large training datasets. Second, we demonstrate a cost-effective training strategy based on simulated measurements, avoiding the effort of assembling extensive experimental databases for cement-based materials. The trained reconstruction network was evaluated on cylindrical specimens composed of an outer, externally exposed ring and an inner, unaffected core. Reconstruction accuracy, quantified by the Structural Similarity Index (SSIM), reached 0.95 for simulated signals and up to 0.82 for real signals. These results indicate that NN-based ultrasonic tomography, trained purely on simulations, can reliably delineate gradient damage fronts and provide a practical tool for non-destructive evaluation of concrete. |
|---|