HIGH PRECISION PREDICTION MODEL OF MATERIAL PROPERTIES BY LOCALLY WEIGHTED REGRESSION METHOD WITH DIMENSION REDUCTION

1 TAKAGI Hiroyuki
Co-authors:
1 KUYAMA Shuji 1 SHIGEMORI Hiroyasu
Institution:
Conference:
27th International Conference on Metallurgy and Materials, Hotel Voronez I, Brno, Czech Republic, EU, May 23rd - 25th 2018
Proceedings:
Proceedings 27th International Conference on Metallurgy and Materials
Pages:
1824-1828
ISBN:
978-80-87294-84-0
ISSN:
2694-9296
Published:
24th October 2018
Proceedings of the conference were published in Web of Science and Scopus.
Metrics:
16 views / 5 downloads
Abstract

To reduce production cost of steel products, it is necessary to control quality stabilization for preventing nonconformance of material properties. The quality stability is improved by constructing a highly accurate prediction model of material properties and incorporating it into the quality control system. In this paper, we aim to establish the highly accurate model of the full-length and full-width prediction of material properties. A locally weighted regression model with dimension reduction has been newly proposed and introduced into the conventional coupled model which consists of a just-in-time statistical model and a metallurgical microstructure model. In the proposed model, the data are projected from the input variable space to a dimension reduction space by a feature extraction method and a regression model with a local weighted method is created from the data near a query point in the projection space. It can properly fit complex process data which have non-linear and co-linear characteristics by mutual complement of these methods. As the result of introducing the proposed model, the standard deviation of an estimated error of tensile strength has decreased remarkably, therefore the target of the estimated accuracy for practical operation has been achieved. The new model has enabled to increase ratio of keeping the material properties within the allowable limits. Hereafter, practical operation by the developed technology has been promoted.

Keywords: quality stabilization; material properties; full-length and full-width prediction; locally weighted regression model with dimension reduction.
Scroll to Top