Digital Twin–Enabled Machine Learning for Smart Manufacturing

Abstract

Abstract Production control problems have been solved using analytical and simulation models, but their applications are limited because of the substantial execution times that they generally require. Applications such as order promising do not allow the time to execute such models. This paper proposes confluence of two Industry 4.0 technologies, machine learning and digital twins, for such time-sensitive applications. Surrogate models developed using machine-learning approaches can provide quick responses, but they need large amounts of data collected from operations over long periods to provide answers with acceptable accuracy. A digital twin of manufacturing operations can be used to drastically reduce the time to generate the data needed by machine-learning approaches. This paper describes the concept for digital twin–enabled machine learning and an implementation for order promising using a hypothetical job shop problem. Data are generated across a range of scenarios by the digital twin of the job shop and analyzed using a Gaussian process regression approach. A meta model is generated that is capable of estimating cycle times based on the current situation in the job shop. The example implementation demonstrates the advantages offered by the approach for production planning and control level applications.