Based on data from process or condition monitoring systems, manufacturing processes are to be optimized using machine learning methods and further controlled self-adaptively in real-time.

Traditionally, manufacturing processes have been organized statically or controlled deterministically as needed. The goal now is to implement self-adaptive control systems for machines and plants to reduce waste and increase efficiency. Using process or condition monitoring systems, characteristic data is recorded, allowing the process to adapt in real-time. While many sensor systems provide indirect information about the process or material, the aim is to implement the most effective monitoring concepts possible. This requires interdisciplinary collaboration between materials and production research and computer science.

Current data streams, combined with access to both current and historical data, are fed into machine learning methods (e.g., adaptive neural networks). This allows for continuous process optimization. Depending on the process, this can be used in real-time for closed-loop control, or the system can be adapted for the next production cycle. This way, deviations in the manufacturing process-caused by factors such as varying material properties or wear of tools, machines, or equipment can be compensated for inline or necessary measures for future processes can be implemented. This not only includes adapting process parameters to deviations but also predictive maintenance methods to ensure quality and optimal utilization of production resources.

Moreover, transferring plant control to an AI-based system allows for incorporating human expert knowledge by including plant settings from human operators and evaluating their effectiveness with regard to production results.

For some processes, especially those with low production volumes or limited accessibility for sensor technology (e.g., high-temperature processes), sensor data acquisition is limited. To build a sufficient database for machine learning methods, experimental data can be supplemented with synthetic data from mathematical models and simulations.