Improving the characterization of porosity by ultrasonic testing in composite materials using X-ray computed tomography, machine learning and signal processing

• Autores: Juan Ignacio Caballero Garzón
• Directores de la Tesis: Federico Sket (codir. tes.) , Ernestina Menasalvas (codir. tes.)
• Lectura: En la Universidad Politécnica de Madrid ( España ) en 2024
• Idioma: inglés
• Tribunal Calificador de la Tesis: Alejandro Rodríguez Gonzalez (presid.) , Carlos Daniel González Martinez (secret.) , Maria Ester Vidal (voc.) , Fernando Adrian Lasagni (voc.) , Sebastián Ventura Soto (voc.)
• Programa de doctorado: Programa de Doctorado en Software, Sistemas y Computación por la Universidad Politécnica de Madrid
• Calificación: Sobresaliente Cum Laude
• Enlaces
  • Tesis en acceso abierto en: Archivo Digital UPM
• Resumen

  • Composite materials are increasingly being used in different industries due to their stiffness and low weight, which makes them attractive to increase efficiency and reduce emissions. Their manufacturing is subject to defects, reason why non-destructive techniques (NDT) have been developed to certify quality standards. Ultrasonic testing, consisting of the propagation of sound waves between 0.5 and 15 MHz, is one of the most widely used techniques because of its low cost and ease of use. Among the different types of defects, the porosity -referred to the appearance of voids in the material- is one of the most common. The state-of-the-art showed a relation between the attenuation of the sound wave and the presence of voids, and this is the main method to detect porosity areas in a component. However, the estimation of the amount of porosity, measured by void volume fraction, and morphological aspects like the void shape, size, and distribution; are not possible to be measured by ultrasonic means. In recent years, machine learning has gained interest in the field of NDT for different use cases, including denoising, improving imaging, automating the detection of defects, and estimation of defects properties. This thesis is developed in this context, particularly in the development of data-science and machine learning techniques for the improvement of the assessment of porosity. For this purpose, the role of X-ray computed tomography (XCT) was fundamental. XCT is a NDT technique that enables 3D reconstruction of composite materials, including their defects. This thesis proposes a data-driven methodology for estimating porosity from XCT and ultrasonic data using machine learning models. Then, thanks to the optimization of the ultrasonic inspection, a methodology and experiments were developed to segment the individual voids by means of convolutional neural networks. Two different methodologies are suggested to classify void morphology solely based on the ultrasonic data. For this purpose, the analytical signal and the contrast between different frequencies was used.