Designing Deep Learning-Based Tools Leveraging Production Data to Support Manufacturability Analysis
R. Siegfried, M. Villamizar, S. Devenes, R. Rey, A. Bacha, B. Morere, SZahno, J-M. Odobez
SSRN, 2024

Abstract- While bars of aluminium profiles are largely used in many industries, their manufacturing through extrusion is challenging. In particular, depending on the complexity of their shapes, significant engineering efforts are involved in determining the feasibility of producing a profile, like selecting the appropriate alloy, determining the appropriate heating temperature, and more generally, the process parameters. To alleviate these challenges, the paper introduces a system with a graphical interface that aims to assist press operators in making informed decisions. The system utilizes two main modules based on deep learning. The first module allows exploration of historical production data to find profiles similar to a given query profile, using geometric and categorical parameters as well as a Convolutional Neural Network (CNN) to compare shape similarity. The second module is a novel Deep Neural Network (DNN) that predicts the required maximum pressure for extruding a profile at a specific speed and temperature. The DNN is designed to capture the physics trend of extrusion with respect to process parameters. A large dataset of production data consisting of more than 1100 clean profile drawings along with the data of 430’000 extrusions was collected and curated for training the deep learning models. Extensive experiments were conducted to validate the trained models and the entire system. The similarity module and its graphical interface were evaluated by experts, demonstrating its effectiveness in quickly identifying revelant profiles for a given query. The pressure prediction module achieved an average error of around 4% in predicting the maximal pressure for extrusion (7.6 or 11.7 bars, depending on the press).

Link

PDF

BibTex