A Study on the Extraction of Geometrical Parameters from Flexible Mechanical Components and Assemblies and their Impact on Performance: A Machine Learning Approach

Over a population of flexible sheet metal assemblies of the same design, such as anautomobile door or hood, there typically are geometric variations that can become apparent as a changing gap between the door/hood and the auto body. The gap variations stem from allowable tolerances for the material (e.g., thickness, strength), spring-back of components post-stamping, and residual stress distortions following clamping and spot welding of the assemblies. To control the geometric effect from these input variations, multi-stage nonlinear finite element analyses (FEA) are used to relate the free-state nodal-point geometry to the inputs. Since the analyses are time-consuming, a sparse grid of input values (the design space) is used. Given the sprung-back nodal data from FEA, the objectives of this research are to construct algorithms that generate geometric performance parameters (e.g. component or assembly twist) and to explore the application of machine learning (ML) for predicting these parameters in flexible assemblies at a greater resolution over the design space than is feasible with FEA. To achieve this goal, three datasets of FEA simulations were examined. Geometric parameters were extracted from two of these. The first was for stamped hat-section sheet metal components and for straight assemblies spot-welded from them, some with a flat unformed sheet as one component. Subsequently, one of the assembly parameters, a minimum-zone magnitude, was used to showcase the suitability of geometric parameters to be used for training a ML algorithm, specifically a fully connected Multi-Layer Perceptron Neural Network. The second dataset consisted of the nodal coordinates from FEA simulations of a Honda-suggested T-joint assembly. Here, algorithms were produced for computing 17 geometric performance parameters. The third dataset comprised multiview images of automobile hood structures. Three deep learning models were compared for accuracy of feature extraction from the images and for directly estimating maximum stress without time-consuming FEA. A last important contribution is the validation of the new iterative virtual-workand- robotics method for fitting nodal points that deviate spatially from a straight line or circular arc. This allows geometric performance parameters for curved components to be included in future flexible assembly datasets.