A1 Vertaisarvioitu alkuperäisartikkeli tieteellisessä lehdessä
Hybrid FE-ML model for turning of 42CrMo4 steel
Tekijät: Laakso, Sampsa Vili Antero; Mityakov, Andrey; Niinimäki, Tom; Ribeiro, Kandice Suane Barros; Bessa, Wallace Moreira
Kustantaja: Elsevier BV
Julkaisuvuosi: 2024
Journal: CIRP Journal of Manufacturing Science and Technology
Tietokannassa oleva lehden nimi: CIRP Journal of Manufacturing Science and Technology
Vuosikerta: 55
Aloitussivu: 333
Lopetussivu: 346
ISSN: 1755-5817
eISSN: 1878-0016
DOI: https://doi.org/10.1016/j.cirpj.2024.10.003
Verkko-osoite: https://doi.org/10.1016/j.cirpj.2024.10.003
Rinnakkaistallenteen osoite: https://research.utu.fi/converis/portal/detail/Publication/458928591
Metal cutting processes contribute significant share of the added value of industrial products. The need for machining has grown exponentially with increasing demands for quality and accuracy, and despite of more than a century of research in the field, there are no reliable and accurate models that describe all the physical phenomena needed to optimize the machining processes. The scientific community has begun to explore hybrid methods instead of expanding the capabilities of individual modelling schemes, which has been more efficient than efficacious direction. Following this trend, we propose a hybrid finite element — machine learning method (FEML) for modelling metal cutting. The advantages of the FEML method are reduced need for experimental data, reduced computational time and improved prediction accuracy. This paper describes the FEML model, which uses a Coupled Eulerian Lagrangian (CEL) formulation and deep neural networks (DNN) from the TensorFlow Python library. The machining experiments include forces, chip morphology and surface roughness. The experimental data was divided into training dataset and validation dataset to confirm the model predictions outside the experimental data range. The hybrid FEML model outperformed the DNN and FEM models independently, by reducing the computational time, improving the average prediction error from 23% to 13% and reduced the need for experimental data by half.
Ladattava julkaisu This is an electronic reprint of the original article. |  |
Julkaisussa olevat rahoitustiedot:
Manufacturing initiative (Susmat) in University of Turku funded by Research Council of Finland (grant nr. 352727).
