Deep Learning-based deposition modelling and path planning in Wire Arc Additive Manufacturing
OPEN ACCESS
Loading...
Author / Producer
Date
2024
Publication Type
Doctoral Thesis
ETH Bibliography
yes
Citations
Altmetric
OPEN ACCESS
Data
Rights / License
Abstract
In Wire Arc Additive Manufacturing (WAAM), the principal goal is to enable the ”first time right“ production of components directly from a digital model, thereby eliminating the need for experimental determination of tool paths and process parameters. Process planning serves as the primary method to achieve this. Tool paths and process parameters are determined through path planning and deposition modeling, which are intended to minimize the reliance on resource intensive experiments. The economic feasibility of WAAM greatly depends on the efficiency and effectiveness of this planning stage. The analysis of the current state of the art in path planning and deposition modelling in WAAM and the related multi-pass welding identified a major limitation; the lack of modelling of actual weld bead geometries considering previous layer geometries. Existing methods predominantly assume flat layers with uniform height. This is due to the absence of a deposition model capable of simulating the weld bead geometry. As a result, physical experiments are still needed for many workpieces. This research addresses these limitations by:
1. Developing a 2D deposition model that accurately represents the cross-sectional shape of weld beads, considering the current geometry of the workpiece.
2. Creating a path planning methodology for 2.5D workpieces that integrates a heuristic approach with layer-wise optimization for layer flatness, utilizing the deposition model to account for variations in layer and bead geometries.
The developed deposition model allows for rapid simulation of weld beads and workpieces in 2D, aiding in the efficient identification of effectivii Abstract ve tool paths. This model can be integrated with existing decomposition and slicing approaches. The model is capable of estimating prediction uncertainty, enable the practitioner to choose adequate safety margins or points for measurement and re-planning. Deposition modeling was decomposed into footprint and shape prediction. A large dataset of 2D cross-sections was collected and used to train neural networks for these predictions. Initially, a transfer learning approach using a pretrained vision network was explored for footprint prediction. However, a Multilayer Perceptron (MLP) architecture was found to be equally as accurate but computationally more efficient. For the shape prediction, predicting the parameters of a B´ezier curve, which approximates the bead shape rather than directly predicting the shape, was found to produce smooth and accurate results. The deposition model was validated through physical experiments and incorporated into a planning method combining a heuristic planning algorithm based on human expertise and layer-wise optimization. This method optimizes each layer for flatness and minimized surplus material. The method was tested in simulation experiments on various workpiece geometries typical for WAAM and multi-pass welding.
Permanent link
Publication status
published
External links
Editor
Contributors
Book title
Journal / series
Volume
Pages / Article No.
Publisher
ETH Zurich
Event
Edition / version
Methods
Software
Geographic location
Date collected
Date created
Subject
Additive Manufacturing; Wire Arc Additive Manufacturing (WAAM); Deep Learning; Deposition model; Wire arc directed energy deposition (DED-Arc); Artificial Intelligence; Machine Learning; WELDING AND ALLIED TECHNIQUES (JOINING OF MATERIALS)
Organisational unit
03641 - Wegener, Konrad (emeritus) / Wegener, Konrad (emeritus)
09706 - Bambach, Markus / Bambach, Markus