AIAx

Initial situation:
Today, the competitiveness of German industry is largely determined by engineering competence and the digitalization of all business processes. The approach and know-how of engineers is strongly individual and can often only be described with high effort or not at all rule-based. The digitalization and networking of devices and work processes increasingly generates heterogeneous data from various sources that are becoming increasingly unmanageable and have not yet been used extensively to optimize business processes. Driven by scientific progress, modern methods of machine learning (ML) have become established in areas such as imaging and speech processing to meet these requirements. A broad applicability in the industry is so far hindered by the lack of traceability of the procedures and their low efficiency with small training data sets (robustness).

Goals:
In AIAx, the existing ML processes will be further developed with regard to traceability and robustness. Reliable results with low data volumes and increased traceability create acceptance, especially among users without an ML background, facilitate knowledge transfer within the company and increase quality. In two applications (Daimler and Endress+Hauser), the advanced ML processes are integrated and tested in assistance systems using the product development process (PEP) in the two key sectors of automotive and electrical engineering.

Approach:
In order to meet the requirement for high robustness, approaches for the fusion of rule- and pattern-based methods are developed in both application cases. Thus, easily formalizable know-how of engineers can already be mapped rule-based and subsequently enriched data-based. From the learned procedure, assistance systems are derived that generate context-based suggestions based on the current development status of a product. In the case of Endress+Hauser applications, the focus is on suggestions for improving a design with regard to its producibility. In the Daimler application, on the other hand, the automated evaluation of simulations using deep learning is focused. In both applications, Active Learning approaches are pursued in order to create acceptance through user involvement and to generate a broader database. Subsequently, metrics for the measurement of robustness and traceability will be developed and the quality of the further developed procedures will be evaluated with the aid of test person studies. Based on existing process models, an action guideline for the use of ML procedures in the engineering domain will be created, which abstracts the project results and best practices from the application cases and makes them accessible to the general public.

Further information can be found here.