Automatic Detection of Geometric Defects in Macroscopic FE Draping Simulation Results
Topic(s) : Manufacturing
Co-authors :
Sophia KELLER (AUSTRIA), Franz MAIER (AUSTRIA), Markus SAUSE (GERMANY), Roland HINTERHÖLZL (AUSTRIA)Abstract :
The development and optimization of a manufacturing process via Finite Element (FE) analysis consists of three main steps: 1) model creation/adaption, 2) process simulation and 3) post-processing. This process can be very computationally expensive and time consuming. There are already tools to enable efficient modifications of the simulation model (e.g. parametrization using Python APIs of simulation software), minimizing the manual effort. However, the post-processing step is a time-consuming task that still requires a lot of expertise and simulation-specific assessment, despite the repetitive nature of the task. The presented study aims to reduce the evaluation effort of draping simulations by establishing an automatic evaluation algorithm.
The draping quality of macroscopic FE simulations is determined by fiber orientation, contour, occurrence of bridging, wrinkles/folds and cracks/gaps. The focus of this study is on the automatic evaluation of the topological defects, i.e. without cracks/gaps. To create a sample dataset for evaluation, we parametrized a draping simulation model using the Python API of the commercial software package Abaqus/CAE. In the simulation model, a fabric blank cut is draped onto a rib geometry using a pressure load on the lamina top surface, inspired by a diaphragm forming step. Load magnitudes across the surface were varied to ensure a variability of draping qualities within the simulation results [1].
Based on those results, we developed a Python script that localizes and categorizes defects. The underlying principle of the algorithm is to compare the simulation result to a defect-free reference mesh of the component.
Each simulation result is analyzed as follows: first, the z-coordinates of the simulation nodes are interpolated at the reference nodes to enable a direct comparison. Then, the deviation values are converted to a binary image using the Yen-thresholding algorithm [2]. In the binary image, a connected components algorithm is used to spot areas with high deviations (Fig.1). The defect areas are mapped back to the point clouds, creating “sub-point-clouds” of the simulation and reference for each defect. For each sub-point-cloud, we fit a polynomial surface and calculate the gradient field of the surface. The differentiation between wrinkles and bridging is done by comparing the linearized gradient fields for each isolated defect using Pearson’s correlation coefficient. Based on the correlation value, the points are categorized. The measurements of the defect are determined by the first three components of a principal component analysis.
Comparisons of manual and automatic defect labeling show that the developed evaluation method works well for the differentiation between bridging and wrinkles if the defects occur at separate locations (Fig.2). In general, the algorithm also works for advanced cases (e.g. transitions from bridging to wrinkling) but is currently being improved for a more reliable distinction.
The draping quality of macroscopic FE simulations is determined by fiber orientation, contour, occurrence of bridging, wrinkles/folds and cracks/gaps. The focus of this study is on the automatic evaluation of the topological defects, i.e. without cracks/gaps. To create a sample dataset for evaluation, we parametrized a draping simulation model using the Python API of the commercial software package Abaqus/CAE. In the simulation model, a fabric blank cut is draped onto a rib geometry using a pressure load on the lamina top surface, inspired by a diaphragm forming step. Load magnitudes across the surface were varied to ensure a variability of draping qualities within the simulation results [1].
Based on those results, we developed a Python script that localizes and categorizes defects. The underlying principle of the algorithm is to compare the simulation result to a defect-free reference mesh of the component.
Each simulation result is analyzed as follows: first, the z-coordinates of the simulation nodes are interpolated at the reference nodes to enable a direct comparison. Then, the deviation values are converted to a binary image using the Yen-thresholding algorithm [2]. In the binary image, a connected components algorithm is used to spot areas with high deviations (Fig.1). The defect areas are mapped back to the point clouds, creating “sub-point-clouds” of the simulation and reference for each defect. For each sub-point-cloud, we fit a polynomial surface and calculate the gradient field of the surface. The differentiation between wrinkles and bridging is done by comparing the linearized gradient fields for each isolated defect using Pearson’s correlation coefficient. Based on the correlation value, the points are categorized. The measurements of the defect are determined by the first three components of a principal component analysis.
Comparisons of manual and automatic defect labeling show that the developed evaluation method works well for the differentiation between bridging and wrinkles if the defects occur at separate locations (Fig.2). In general, the algorithm also works for advanced cases (e.g. transitions from bridging to wrinkling) but is currently being improved for a more reliable distinction.