Initial Situation:

In the context of manufacturing, 3D models play a crucial role. By creating precise and detailed 3D models of manufacturing components, important information can be obtained that enables better analysis and optimization of manufacturing processes. The use of 3D models allows the geometry, structure and properties of the components to be captured and taken into account in the analyses. Three-dimensional deep learning is a promising approach for gaining valuable insights from these 3D models. By using neural networks that have been specially developed for processing 3D data, complex relationships and patterns in the production components can be recognized. These findings can be used for defect detection, quality control, process optimization and the prediction of manufacturing properties.

Possible tasks:

Within the framework of the research project, there are several possible main tasks, each consisting of a structured literature research, conceptual design and implementation. Only one of the following topics will be considered as part of a student project:

• Graph representation learning
• Diffusion-based networks
• Application of heat map regression to triangle meshes and graph data
• Semi- and self-supervised learning approach applied to the below mentioned dataset
• Implementation of a library in Rust (Python bindings) using WebGPU to render and label triangle meshes and point clouds.

The data set for all tasks is already available and can be extended as required. The data is available as a triangle mesh and can also be viewed comparatively as a point cloud. Access to sufficiently large graphics cards is guaranteed. Further information as well as the start and scope of the work can be discussed in a personal meeting. The topic will be specified once personal interests have been discussed. All focal points can be worked on remotely. During the Master’s thesis, students will work with the Electrical and Electronic Components Dataset (Scheffler, Benedikt; Bründl, Patrick, 2023), which is freely accessible in the Harvard Dataverse.

Notes on application:

• Interest in machine learning, deep learning, statistics etc.
• Strong math background beneficial
• Practical online courses and books will be provided after consultation on the existing level of knowledge
• Experience with libraries such as PyTorch, PyTorch Geometric, NumPy, Open3D, etc. required
• IT affinity and experience in Python and at least one of the following programming languages required (e.g., Rust, C++, C, Go)
• Written and spoken German or English required
• The thesis has to be written in English in LaTeX (e.g., TexStudio, Overleaf)
• Literature management must be done using JabRef
• High motivation, curiosity and a structured way of working
• Please send applications with CV and current overview of subjects by e-mail to benedikt.scheffler@faps.fau.de.
• Generic e-mails will be ignored (how to write a proper e-mail).
• In the first meeting there are questions regarding the stated requirements, as well as the below listed material. Based on this, the student’s suitability for this thesis is determined.

Before applying, familiarize yourself with following topics:

• Papers:
  • Electrical and Electronic Components Dataset
  • A Dataset of Electrical Components for Mesh Segmentation and Computational Geometry Research
  • Semantic part segmentation of spatial features via geometric deep learning for automated control cabinet assembly
  • Geometric Deep Learning as an Enabler for Data Consistency and Interoperability in Manufacturing
• Software:
  • Documentation: PyTorch Geometric
  • YouTube: PyTorch Geometric
• Theory:
  • Geometric Deep Learning
  • YouTube: Geometric Deep Learning
  • YouTube: Machine Learning with Graphs

Pure literature research is not possible, as it is always part of practical work as well.

Kategorien:

Kontakt: