Modern production plants are often so complex that the interrelationships can only be incompletely captured by classic modeling. Optimization potential can then only be exploited with data support using machine learning (ML) methods. ML is therefore increasingly being used to increase product quality, reduce the use of resources or avoid unplanned machine downtime through predictive maintenance.
However, the opportunities are offset by major challenges: There is a lack of experts who are equally at home with ML methods and production and automation technology. Reusable components for ML-based systems in the production environment are in short supply. There is no established procedure for large, heterogeneous project teams, and adaptation to changing conditions during operation - wear and tear, properties of input materials, structural modifications to the process - must be guaranteed.

The process model

Six Fraunhofer Institutes have now jointly developed a standardized process model and the associated tools for the use of ML in production as part of the lead project "ML4P - Machine Learning for Production". The AI engineering approach is based on systems engineering in many respects. A chain of interoperable software solutions has been developed. The tools are used to systematically record and formalize the relevant knowledge and data of a production plant and prepare it for the use of the ML method spectrum. Furthermore, they can detect and evaluate existing optimization potential, select the most suitable ML methods for specific applications and use them to their advantage.
"Although the pure AI algorithms are of central importance for the use of ML in production, they often only make up a fraction of the overall solution," explains Dr. Julius Pfrommer, research group leader at Fraunhofer IOSB and team leader of the process model. Another important piece of the puzzle is provided by the process model, which can be widely used regardless of the specific application. It is divided into six phases with clearly defined results and uses two central documents or data structures that represent the current state of knowledge across all phases: the machine learning pipeline diagram and the virtual process file.

AI algorithms - only a fraction of the overall solution

"Although the pure AI algorithms are of central importance for the use of ML in production, they often only make up a fraction of the overall solution," explains Pfrommer. Another important piece of the puzzle is provided by the process model, which can be widely used regardless of the specific application. It is divided into six phases with clearly defined results and uses two central documents or data structures that represent the current state of knowledge across all phases: the machine learning pipeline diagram and the virtual process file.

Integrating expert knowledge in a targeted manner

There is also a role model that covers the disciplines, skills and functions required in each phase. Pfrommer: "This comprehensively describes the path from problem definition to continuous operation of the ML-based system. In particular, it defines the knowledge management and interfaces required to enable scaling to large teams."
An important aspect is the targeted integration of the specifics and prior knowledge from the application domain, the researcher continues. "The expert knowledge from the engineering disciplines is a great treasure. You can't simply superimpose a neural network onto the existing models. Instead, a deep integration of existing tools from engineering disciplines with the AI processes must be achieved. This is the only way to ensure that AI can also do a good job in areas where it has little or no data and empirical values from the past."
In addition to the Karlsruhe and Lemgo sites of the lead Fraunhofer IOSB, the Fraunhofer Institutes for Intelligent Analysis and Information Systems IAIS, for Factory Operation and Automation IFF, for Industrial Mathematics ITWM, for Mechanics of Materials IWM and for Machine Tools and Forming Technology IWU are also involved in the work on the process model.
The ML4P White Paper will be presented to the public for the first time on October 26, 2020. The white paper will then be available for download from Fraunhofer IOSB.