Technical products are becoming more independent, flexible and capable of action due to the increasing integration of information and automation technology. However, due to the increased integration of functions, systems are becoming more complex and personnel are finding it steadily more difficult to estimate the extent of their actions or to identify necessary actions. Both problem areas have in common that a prediction of the machine behavior is necessary. In many cases, attempts are made to circumvent this with the help of massive computing power in the cloud, but for many data, a transfer to the cloud is not technically and economically feasible due to too high a resolution, and is also not desired due to data protection. The solution to this is so-called edge computing, which is intended to combine the advantages of modern algorithms, such as machine learning, with the machine proximity of a PLC.

In this project, decentralized Machine Learning (ML) algorithms for embedded and edge devices will be developed. The selection and use of ML algorithms depends essentially on the goal, the available information and thus on the respective application. A major goal of this project is therefore the development and implementation of learning methods in distributed, embedded systems under consideration of resource constraints. The starting point for this is data obtained from multimodal, active sensor systems and evaluated using embedded, real-time data analysis methods as well as ML methods. These data are to be further processed locally in the machine or component and thus a data-centric automation of machine and plant engineering is to be developed.

The main goals of the project are the analysis of ML methods considering expert knowledge, interpretability of learning results, embedded learning on data streams from active sensor systems in real-time and their transferability into practical applications. In addition, the development of runtime-optimized ML algorithms for embedded and edge devices, will also be worked up for practice-relevant findings. In the application context, the transfer of ML methods into three concrete use cases will take place: Predictive Quality in Wrapping Technology, Predictive Maintenance for Drive and Control Solutions in Conveyor Applications, and 3D Vision in Robotics. The use cases considered in this project are intelligent quality prediction of profile wrapping machines, predictive maintenance of failure-critical components in conveyor applications, such as slippage of friction wheels and drive belts due to wear or damage to bearings and gears, and holistic production planning. These use cases are at the machine or component level. Therefore, the deployment of ML has to be done on the limited resources of the component (embedded device) or the machine (edge device) in order to integrate the ML technologies into the products and into the production machine. To achieve the goals, defined work packages are processed in the use cases, from the requirements analysis to the selection of learning methods with integration of expert knowledge, to learning on data streams and runtime optimization of the algorithms for edge and embedded devices. To achieve the goals, a regular transfer of expertise and experience between the project partners will be carried out during the project duration. In particular, the international exchange between German and Canadian partners aims to bring together the competencies of the project partners.

The project is carried out as a joint project between the Canadian company NTWIST (Industrial AI), the Canadian research partner ACIS (Advanced Control & Intelligent Systems Lab of the University of Victoria), the German companies düspohl and encoway, and the research partner Fraunhofer IEM. The specific know-how of the companies düspohl and encoway will be complemented in particular by the methods and experience with self-optimizing systems and machine learning of the IEM. Following the 3-year funding period, the immediate economic exploitation of the quality prediction for the RoboWrap and the predictive maintenance for the electric drive technology is aimed at in order to equip the high-wage production locations Germany and Canada for the future. The objective of EASY shows high practical relevance and a clear innovation potential especially for SMEs. At the same time, however, the aim is to make the results scalable so that large companies can also use them. Thus, a sustainable positive impact on the German and Canadian economy, especially in terms of maintaining their own position in the international market, should be achieved.