AI-Based Optimization in Automobile Production

Traditionally, automotive production already has a high level of automation (around 1,000 robots per 10,000 workers), the largest share of which is attributable to industrial robotics with classic assembly line systems under controllable conditions. However, global developments such as climate change or megatrends such as increasingly specialized customer requirements are imposing new challenges: Alternate fuels, lightweight materials that are more complex to process, and greater flexibility in production are just a few of them. Conversely, software tools from artificial intelligence are reaching an ever higher level of maturity and are already helping various (primarily still digital) industries to achieve leap innovations. Especially in combination with meaningful data volumes, numerous routine activities that were previously only humanly feasible can be supported or even automated in order to produce more efficiently and conserve resources. Exemplary applications can be found in the prediction of scrap, automated fault diagnosis or in the optimized control of production processes. Furthermore, combining learning algorithms together with engineering simulations can lead to an improved understanding of processes.

Research in this competence area therefore investigates both application potentials of modern deep learning systems such as CNNs, GANs or GNNs as well as probabilistic learning methods such as Bayes nets, and combinatorial optimization (CP/MIP), in particular:

Intelligent process monitoring in manufacturing
•    Prediction of scrap or quality results with suitable sensor technology, e.g. in resin injection processes, machining or additive manufacturing;
•    Predictive maintenance for minimizing downtimes and root cause analysis.

Digital twins
•    Combination of conventional engineering models (FEM, PDE) with machine learning with feedback of real data for process control;
•    Deep learning in physical simulations (e.g., with neural networks for graphs, GNNs);
•    Transfer learning from simulated to real process data;
•    Hybrid models: constrained deep learning / probabilistic machine learning (prior knowledge + learning).

Optimization and decision making
•    Algorithms and software for discrete optimization: task and resource allocation, packing problems, scheduling; constraint programming with MiniZinc (e.g. Google OR tools);
•    Fair optimization e.g. for shared/cloud manufacturing;
•    Reinforcement learning.

The considered topics are addressed in teaching accompanying the research:
•    Elective courses in the field of AI;
•    Introduction to AI and Neural Networks (Master APE);
•    Practical seminars and practical courses;
•    Theses (Bachelor/Master);
•    Cooperative PhD projects.