Shape parameter estimation of a 3D model from multi-view radiographic images

High pressure turbine blades in aircraft engines are exposed to extreme thermal and mechanical conditions.They withstand them through multiple innovations, which demand strict controls to be carried out during their production.Among the Non Destructive Evaluation (NDE) methods, X-ray radiography offers the ability to image the internal (complex) structure of the blade.To meet the production requirements, this control must be fast, which translates into the acquisition and analysis of a limited number of radiographic images of the blade, typically around ten.In this context, the thesis work aims at developing a NDE method based on a limited number of radiographic images, for the identification and characterization of geometrical indications.The selected strategy to address this challenge consists in generating a deformable 3D model of the blade registered onto the observed images using a Projection-based Digital Volume Correlation (P-DVC) approach.The dimensions to be checked are calculated on the registered model.The P-DVC relies on the minimization of a cost function involving the projection residuals, defined as the differences between observed and numerically simulated images.First, a study of the noise present in radiographic images has been conducted.The properties of the noise polluting a flat-field image were first determined (normality, variance, spatial correlations), from which those of the noise of a radiographic image were inferred and verified experimentally.This provides a reference measurement to interpret the projection residuals (noise level), and to introduce an cost function optimal for the acquisition noise.Additionally, the physical phenomena involved in the formation of radiographic images must be identified and correctly calibrated to simulate realistic radiographic images, comparable to those observed.Beam Hardening and Compton scattering are the two main physical mechanisms that impede quantitative image analysis.Parametric models to reproduce these phenomena on simulated images have been proposed.The associated parameters are iteratively identified through P-DVC.The projection residuals obtained using the nominal model of the blade (i.e. its CAD model) are processed to identify geometric indications.Their characterization is achieved through the deformation of this model, so as to generate the above-mentioned corrected model.This deformation implies the design of a deformation model, which needs to be rich enough to represent the expected shape variability while containing a reasonable (i.e. small) number of parameters to ensure a good conditionning, and hence a low uncertainty.A model derived from the manufacturing process of the blade has been designed.It consists in the partition of the part into multiple subparts: one representing the external surface, five representing the different elements of the ceramic core describing the intern geometry.Their relative kinematics is given by a rigid-body motion, plus a scale factor to account for thermal shrinkage.The wall thickness measurements of the blade calculated on the registered model match well those calculated on a tomographic volume.Further reflections for improving the precision of the measurements are considered, with a view to finding an alternative economically more interesting than tomography and richer than the ultrasound controls currently carried out on the blades.Additionally, the generic nature of the method allows its application to the NDE of other metallic parts.