Evaluation of high-dimensional measurement data from industrial manufacturing for process monitoring and quality prediction

Evaluation of high-dimensional measurement data from industrial manufacturing for process monitoring and quality prediction

Published November 2019.

Downtimes of machines and systems in manufacturing are expensive and should be avoided as far as possible. Process monitoring methods are suitable for detecting deviations from regular plant operation. As a result, malfunctions can be detected and remedied at an early stage. The challenge here lies in identifying the position of the fault in the process.

In addition, quality predictions based on the set process parameters can prevent the production of rejects. Often, the quality of the manufactured product can only be determined in elaborate laboratory measurements, so that this information is available with a time delay. An immediate prediction after measuring the process parameters saves valuable time here.

The contribution of this project is the evaluation of the data to address the above mentioned problems. The unfunded project partner Dieffenbacher provided data from its cooperating customers, which were first used to test and validate ML methods for their suitability with respect to the requirements in the context of process monitoring and quality prediction. Suitable methods were selected, and their usefulness in application was demonstrated prototypically.

In the first step, high-dimensional data were processed in close cooperation with process experts to ensure their usability for machine learning methods and the information content. In this process, features were aggregated in a suitable way to reduce the dimensions of the data set.

In the second step, the processed data were used for continuous monitoring of the manufacturing process, with machine learning (ML) methods trained to detect regular operating states of the equipment. Deviations from regular operation can thus be detected by the algorithm and communicated to the operator.

In the second step, product quality data measured in the laboratory were included in addition to the process data. It was shown that the process data contain significant information about the expected quality, so that quality predictions can be derived from the process data.

The exploitation of the scientific results is carried out by Fraunhofer IOSB. Processes and algorithms will be used in further research and industrial projects. A publication is planned for CASE 2021 (IEEE International Conference on Automation Science and Engineering). In addition, the project partner Dieffenbacher and Fraunhofer IOSB plan to jointly patent the developed ideas and results.

Dieffenbacher is active in the field of mechanical and plant engineering and manufactures press systems, as well as complete production lines for the wood, automotive, aerospace and recycling industries. Dieffenbacher employs over 1,600 people and operates at 16 production, service and sales locations worldwide.

Project period

01.11.2019 – 30.06.2020

Contact

Dr.-Ing. Julius Pfrommer (julius.pfrommer@iosb.fraunhofer.de), Fraunhofer IOSB-ILT

Dr.-Ing. Thomas Usländer (thomas.uslaender@iosb.fraunhofer.de) , Fraunhofer IOSB-ILT

Jürgen Woll (juergen.woll@dieffenbacher.de), Dieffenbacher GMBH Maschinen- und Anlagenbau

Dr. rer. nat. Constanze Hasterok (constanze.hasterok@iosb.fraunhofer.de), Fraunhofer IOSB-ILT

Jürgen Woll (juergen.woll@dieffenbacher.de), Dieffenbacher GMBH Maschinen- und Anlagenbau