Systems based on conventional image processing are often used to ensure the quality of production steps. Programming them is complex and their use - especially with a low cycle time or a high production rate - reaches the limits of computing power and reliability.
Through the use of deep learning or neural networks, some quality assurance tasks can be solved more effectively and more stably than with conventional image processing - in some cases, even tasks that cannot be implemented conventionally can be solved. The large number of required classified training images - for example images with classified error groups - turns out to be a hurdle here. Since their acquisition is associated with great effort, the use of deep learning is only worthwhile for products with large quantities - use for small batch sizes is currently not economical.
Aim of this project in cooperation with ETEC - Automatisierungstechnik Ges.m.b.H. is the creation of a method for the generation of faulty classified training images from CAD data of products including evaluation of the limits and feasibility. Suitable deep learning networks are then identified, trained and their accuracy evaluated. By combining the results in a test facility and applying them to different product families, decision and success factors for quality assurance measures based on deep learning are identified and made usable in a decision matrix.
For more information on the project, please contact:
Benjamin Massow Lecturer Mechatronics & Smart Technologies +43 512 2070 – 3924 email@example.com
Foto: Adobe Stock - xiaoliangge