Novel metrology tools based on artificial intelligence to study damage and fracture in aerospace structures

Abstract

This PhD thesis adapts the technology developed in raising fields in computer science, such as Machine Learning and Computer Vision, in order to create novel metrology tools that enable the qualitative and quantitative assessment of damage and fracture in Aerospace structures. The developed tools are easily integrated in the industrial functions and replace manual human-based processes with computerized automated operations. These new metrology methods enable efficient inspections and material failure analysis in two different scales: In macroscopic level: the development of a new metrology system that enables inspections in large scale aerospace components is presented. The main purpose is the design and assembly of a metrology system that performs inspections in specific locations along the large surface of an aerospace component and produce accurate digital representations of the inspected areas. Additionally, a metrology framework that allows this fully operational metrology system to perform inspections in targeted positions is established. The functionality of this tool is evaluated in a collaborative project between the SME that hosted this research work, AEROSERTEC, and AIRBUS, where the objective was the improvement of the repairing process of the carbon composite aerospace components. The capability of the produced metrology to create accurate 3D digital representations of the inspected areas is implemented in one of the intermediate steps of the repairing process and it is shown that it can contribute to the optimization of the process. In microscopic level: the objective is to introduce methods already mature in Computer Science and Artificial Intelligence in order to create tools that can perform topographic characterization of the fracture surface of engineering materials. The developed Machine Learning algorithms are divided in supervised, where a pre-trained model is necessary for performing predictions, and unsupervised, where the algorithm can perform predictions without prior training. The supervised methods are implemented in the topographic characterization of the material's fracture surface and enable accurate measurement of the relative appearance of the different micro-structural fracture modes in the microscopy images. The unsupervised methods are used to cluster fracture surfaces of different samples produced under different experimental conditions and allow their classification according to external parameters (imposed loading, environmental conditions etc.) and internal micro-structure characteristics. The evaluation of their results reveals great potential, and the comparison with current methods that perform fracture surface classification, shows superior efficiency.