Companies such as the Miba-Group which manufacture a wide variety of products by using an enormous range of production methods and processes are facing a particular challenge in the current trend towards flexible production in small lot sizes and with constantly changing and extended product type ranges. In such an environment, the generally extremely high quality requirements must also meet economical aspects at all time.
Transfer-Learning based Framework for automated Quality Inspection
This challenge is (in the case of Miba) additionally exacerbated by the global distribution of production sites with varying quality standards and conditions. Conventional quality inspection systems reach their limits, as these systems only use the direct task specific data for the respective inspection problem.
This project addresses the above mentioned problem by using deep transfer learning methods in order to transfer quality-related knowledge from a part A to a part B. The objective here is to optimize quality inspection models by utilizing so far unused data with different modalities from different inspection systems. To ensure the generality of the approach, the framework will be developed by using data with different modalities (optical, acoustical etc.) from different use cases. Finally, the methods will be evaluated, especially with small lot sizes, on production-variants of sinter metal parts, friction discs and engine bearings for the automotive industry.
The central result of the project will be a software framework designed as a “fog computing cloud” for data-driven methodologies and workflows for a later use in the industrial environment.
The project demonstrates on the basis of automated, non-destructive quality inspection systems, how effective data-driven modeling techniques of artificial intelligence can be adapted using transfer learning to meet the requirements of the manufacturing industry and to exploit synergies. In this way, this project will provide important conceptual and technical foundations for the upcoming transformation process of digitization and flexibilization.
This research project is a cooperation between experts of the Miba-Group (quality expertise in the automotive industry, use cases and data science infrastructure), RECENDT (sensor technology, process integrated inspection) and SCCH (image processing, deep and transfer learning, software framework).