Contribution à l’inspection des pièces mécaniques souples sans le recours à des opérations de conformation

Abstract

The dimensional inspection of manufactured parts is an essential step during the manufacture and assembly of components. Due to inherent variations in manufacturing processes, produced components contain dimensional and geometric deviations from their nominal shape as defined in computer-aided design (CAD) software. These deviations are referred to as "defects". Note that the defects may or may not be acceptable, depending on the tolerance allowed. In addition, and in the specific case of non-rigid (or flexible) parts, the shape (geometry) in the Free State can be significantly different from the nominal shape. These "deformations" are caused by gravity and / or by the distortion induced by residual stresses. Traditionally, the separation of deformations and defects is achieved by conforming the component to a jig which reproduces the nominal shape. Therefore, the defect inspection is carried out on a template specifically dedicated to this task. If we want to proceed the inspection without a conforming device, mathematical manipulation becomes necessary to virtually conform the part based on registration operation called non-rigid. Therefore, the goal of this research is to provide a digital tool to manipulate a point cloud resulting from a digitization without using special conforming template for a flexible mechanical component. Thus, inspection process decreases costs and increases productivity. More specifically, the project aims to improve CPD algorithm (Coherent Point Drift) widely used in imaging applications and not adapted for mechanical components. Our work improves the search for correspondence between the measured points cloud and a nominal point cloud by performing a flexible registration. In other words, the developed registration operation guarantees an isometric transformation restricted constraints reflecting the physical properties of the mechanical part (tensile rigidity, torsional flexibility). The main idea of our work is that during the alignment phase with the CPD algorithm (iteratively executed to align the CAD meshes on the SCAN meshes), two criteria are simultaneously minimized: the point to point distance between CAD and SCAN and a scalar representation of change of parameter size of CAD mesh at each iteration of CPD algorithm. The proposed project is based on two main contributions. The first is to improve the alignment process by developing an optimization method based on genetic algorithm (GA) to find the optimal regularization parameters of CPD algorithm that control the alignment phase. Also, we have adapted the hyper-parameters Of CPD to the local stiffness. This approach is effective especially in the case of those representing rigidity variations located in regions with strong curvature. A reformulation of the objective function of the algorithm is put forward. A correction matrix associated with the importance of each measurement point, and therefore with the local stiffness of the part during the alignment phase, has been introduced and integrated. The measurement points representing the same stiffness were grouped and classified using the fuzzy c-means algorithm. The effectiveness of the proposed approach is demonstrated on various case studies from the transport industry. The results obtained demonstrate better detection and quantification of dimensional and geometric errors. The second contribution of this thesis project is to propose another approach complementary to the first contribution to optimize the alignment phase. This is an improvement of the Euclidean distance calculation criterion between the two CAD and SCAN models by defining metrics based on a distance calculation approximation in order to obtain a good estimate of defects.

Date	31 Aug 2021
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Antoine Tahan (Supervisor) & Mohamed Ali Mahjoub (Co-supervisor)