Predicting Production and Assembly Times in Steel Manufacturing Using Machine Learning

Abstract

The steel manufacturing industry is undergoing a gradual transition towards the integration of digital processes throughout its entire supply chain. New technologies, such as artificial intelligence, have been the subject of greater exposure, and new possibilities have been identified for enhancing production planning.

In collaboration with the VVTW company, this study seeks to develop data-driven models for the precise prediction of production and assembly times for a range of steel components, including positions and assemblies.

The initial step involved the extraction of meaningful data from the company’s comprehensive database. Following a thorough examination of the data and the extraction of valuable insights for the company, three distinct models were refined and evaluated in comparison to a representative baseline currently employed in steel factories.

With regard to the evaluation of positions, the Random Forest model exhibited the most favorable performance, demonstrating a 71.05% and 74.93% improvement of RMSE over the baseline, respectively, for plates and profiles. Concerning assemblies, the most effective model was also the Random Forest, which yielded an improvement of 60.04% in RMSPE relative to the baseline. Additionally, a particular category of low-weight assemblies appeared to be more challenging to accurately predict. To address this, two potential explanations were put forth.

Ultimately, the data-driven models were incorporated into the company’s software platform for implementation with other clientele.