Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

Integration of AI based data analytics into industrial applications

Presentation held at EPoSS Annual Forum 2021, 04-07 October 2021, Freiburg im Breisgau, Germany
: Mayer, Dirk

Präsentation urn:nbn:de:0011-n-6425394 (1.7 MByte PDF)
MD5 Fingerprint: af9a6a54e91efcbfcd789bee6375630b
Erstellt am: 28.10.2021

2021, 14 Folien
European Technology Platform on Smart Systems Integration (EPoSS Annual Forum) <2021, Freiburg/Brsg.>
Vortrag, Elektronische Publikation
Fraunhofer IIS, Institutsteil Entwicklung Adaptiver Systeme (EAS) ()

Since a decade, digital technologies are integrated into industrial processes and assets with the aim to increase flexibility (e.g. for mass customization) and efficiency. Thus, industrial environments procure large data sets on different levels from the ERP system, over factory automation to sensor devices. Machine learning and other methods of artificial intelligence enable data driven modelling, analysis and optimization beyond traditional methods of process identification and control. However, in contrast to other AI applications, industrial processes have some specific requirements. The data is collected in industrial automation networks; For closed loop control, latency should be kept low; industrial users partially prefer sensitive data not to be processed in remote cloud servers; Bandwidth of (wireless) networks is limited, so transmission of large data amounts in real time is rather impossible. A possible solution is shifting the data analysis from a central instance towards the edge of the network, or even to smart sensors. This leads to new tasks for the design of AI systems including a co-design of hardware and software, sometimes also including communication systems. The distribution of AI in a sensor-edge-cloud network has to be implemented considering the data acquisition, training and inference stages of operation. For instance, training of a complex machine learning algorithm might be implemented on a powerful central server, while the trained network can be implemented on an IoT device. Also the target hardware has to be chosen appropriately: A dedicated hardware with low flexibility might be produced at very low cost per unit, while a reconfigurable or programmable platform would be easier adapted, but would be less performant regarding costs, size and energy consumption. In this contribution examples from current research projects on industrial applications of AI are reported with a focus on the question of systems integration with sensor-edge-cloud systems. The discussion also includes also aspects of model based engineering of edge and networked systems.