Advancing Incremental Profile Forming: Experimental Analysis of Axial Grooving and Deep Learning Modeling
Loading...
Date
2024-05
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
The Ohio State University
Abstract
This thesis presents the experimental analysis of the axial grooving process during Incremental Profile Forming (IPF), a manufacturing process developed at TU Dortmund, Germany [1]. The IPF machine has advanced machine control but limited capability to accommodate certain process mechanics like springback, resulting in inaccurate part geometry [2]. A computationally efficient model of axial grooving is sought that can be used for process control. The goals of this thesis are to obtain a better understanding of the axial grooving process and to investigate the efficacy of modeling deformation using Deep Learning (DL), which has been shown to be an efficient modeling tool [3].
Axial grooving processes were performed on the IPF machine, and geometry data collected. The experimental process was then simulated using the finite element method (FEM), and resulting geometry data was compared to measured data to gain insight on the nature of deformation, the accuracy of the FEM model, and the accuracy of the IPF sensors.
Experimental results show that laser line optical sensors are accurate to within the desired error threshold of 0.1 mm, however the current coordinate transformation and alignment procedure results in an error of 0.18 mm. Geometric data shows uneven groove depth between the entry, steady-state, and exit regions of the grooves. Tool type was shown to have an effect on overall groove geometry.
For the DL investigation, a 2D beam bending problem [4] was modeled using both data driven neural networks and Physics Informed Neural Networks (PINNs). A study was done to understand the effect of input normalization on model accuracy. The PINN predictions were compared to data-driven network model outputs and the known analytical solutions. A second study was performed to evaluate the effect of input domain size on PINN model accuracy.
It was determined that input normalization allows data-driven models to handle vast input domains but introduces error in PINN models. From the domain study, it was shown that PINNs are more accurate if their input domains are restricted.
This research will advance understanding of IPF process mechanics including errors involved in on-line sensing. Furthermore, the DL study has identified key limitations of using PINNs for modeling problems with large input and output domains.
Description
Keywords
incremental profile forming, deep learning, metal forming, physics informed neural networks