Data-driven modelling of multi-stage stamping and bending processes

Project Overview

Principle Investigators:	Prof. Dr.-Ing. Werner Homberg Prof. Dr. Barbara Hammer Prof. Dr.-Ing. habil. Ansgar Trächtler
Project Team:	Andreas Mazur M.Sc. Henning Peters M.Sc.
Research Institutions:	Institute of Forming and Machining Technology (LUF), Paderborn UniversityCentre for Cognitive Interaction Technology (CITEC), Bielefeld University Fraunhofer Institute for Mechatronic Design Technology (IEM), Paderborn Heinz Nixdorf Institute (HNI), Paderborn University
Semi-finished Material(s):	Flat steel wire (1.4310)
Manufacturing Processes:	Roller levelling, stamping and bending

Motivation

Consideration of cross-stage and quantity-dependent effects
Development of a hybrid, data-driven process modelling for a multi-stage straightening and punching-bending process for flat wire
Development of a model-based process control system

Aims

One challenge of the straightening and stamping and bending process with regard to repeatable component quality is the cross-stage and quantity-dependent effects that are induced in the wire due to the wire production and the winding of the wire onto coils. The aim of the project is the data-driven modelling of a multi-stage and quantity-dependent straightening and stamping-bending process in order to develop a model-based process control. This process consists of a flat wire coil, a roller leveller and a stamping and bending machine. With regard to process modelling, a hybrid approach is being pursued using a combination of data-driven methods and knowledge-based models. The hybrid machine learning module (HML module) to be developed in this way consists of a digital twin that takes the data from numerical or analytical models as well as expert knowledge and sensor data from the real process, processes it and then transfers it to the machine learning module (ML module). The aim of the ML module is to learn compact representations for high-dimensional process data, on the basis of which a surrogate model is to be trained. The task of the surrogate model is to efficiently map complex and non-linear correlations between cross-stage and quantity-dependent process data, which can currently only be estimated approximately using expert knowledge. Process corrections should be estimated using the efficient surrogate model and sent back to the digital twin. The digital twin checks the proposed actuator adjustments and, if necessary, forwards them back to the actuators of the stamping and bending section.

Working Program

Work Package	Description
WP1	Research into the acquisition and formalisation of relevant forming system data using suitable mechatronic systems
WP2	Formalisation of the processes under consideration, automated data acquisition and specific domain knowledge
WP3	Implementation of the hybrid machine learning module
WP4	Provision of reference data sets and integrated application scenarios as well as validation of research results
WP5	Documentation in the sense of an evaluation of the data-driven straightening and punching-bending process with regard to the developed HML module

Expected Results

The expected result of the research project is the modeling of a straightening and stamping-bending process for the correction of cross-stage and quantity-dependent effects and, consequently, an increase in product quality. For the hybrid modeling approach, analytical and numerical approaches as well as experimental investigations and expert knowledge are combined with machine learning methods. This results in the HML module consisting of the digital twin and the ML module.