Closed-Loop-Regelung eines virtuellen Umformprozesses

• Unité de recherche: COST
• Numéro de projet: C00.0081
• Titre du projet anglais: Closed-Loop-Regelung eines virtuellen Umformprozesses

Mots-clé (Anglais)

Mdelling; process optimization; neural networks

Programme de recherche (Anglais)

COST-Action P4 - Non-linear dynamics in mechanical processing

Description succincte (Anglais)

See abstract

Autres indications (Anglais)

Full name of research-institution/enterprise: Fachhochschule für Technik St. Gallen

Partenaires et organisations internationales (Anglais)

CZ, DK, F, D, GR, H, IRL, I, NL, PL, SI, E, S, CH, GB

Résumé des résultats (Abstract) (Anglais)

To optimise a production process is a difficult, expensive, time-consuming and sometimes dangerous procedure - especially if more than 2 or 3 process parameters are involved or if the process is non-linear. In practical applications this procedure is often carried out on a trial-and-error basis. More systematic Design of Experiment methods do exist, but the practitioner most often considers them as too expensive and too rigid. In this project the optimisation of a metal forming process is considered in order to investigate the optimisation methodology. A cylindrical piece of steel is is deformed in a press, with length, diameter and thickness of the deformed piece as product properties. The crucial question in process optimisation is: Which process parameters lead to a product with the desired product properties? This question should be answered with as few experiments as possible. On order to minimize costs, a first optimisation phase is carried out virtually. The production process is substituted by a Finite Element calculation. Even on a powerful computer a single FE-calculation takes several hours. Since the optimisation requires several hundred calculations, a less time consuming model is required. Adaptive Neuro-Fuzzy Inference Systems, trained with the results of a number of FEM-calculations, turned out to mimic the FE calculation successfully. Since a recall of a small neuro-fuzzy inference system takes in the order of milliseconds these systems are apt for integration in a real-time optimisation loop. It could be shown that with 3 input dimensions 30 virtual experiments are sufficient to find the desired operating point - significantly less than with classical Design of Experiments methods.

Références bases de données (Anglais)

Swiss Database: COST-DB of the State Secretariat for Education and Research Hallwylstrasse 4 CH-3003 Berne, Switzerland Tel. [phone] Swiss Project-Number: C00.0081