Digital twin-driven supervised machine learning for the development of artificial intelligence applications in manufacturing

TLDR

Digital Twins link physical and cyber spaces and AI‑based machine learning is widely seen as promising for manufacturing, yet supervised ML requires large, manually labeled datasets that are costly, error‑prone, and time‑intensive in complex production settings. The study proposes using Digital Twin models to accelerate ML training by automatically generating and labeling synthetic datasets, thereby reducing user involvement. The framework creates synthetic data from DT simulations, augments it with limited real‑world information for cross‑validation, and automatically labels the data for ML model development. In an industrial use case, the framework successfully trained vision‑based part‑orientation recognition from DT‑generated data, enabling adaptive control of the production process.

Abstract

Digital Twin (DT) implementations can contribute to smart manufacturing by integrating the physical and the cyber space. Artificial Intelligence (AI) applications based on Machine Learning (ML) are widely accepted as promising technologies in manufacturing. However, ML methods require large volumes of quality training datasets and in the case of supervised ML, manual input is usually required for labelling those datasets. Such an approach is expensive, prone to errors and labour as well as time-intensive, especially in a highly complex and dynamic production environment. DT models can be utilized for accelerating the training phase in ML by creating suitable training datasets as well as by automatic labelling via the simulation tools chain and thus alleviating user’s involvement during training. These synthetic datasets can be enhanced and cross-validated with real-world information which is not required to be extensive. A framework for implementing the proposed DT-driven approach for developing ML models is presented. The proposed framework has been implemented in an industrially relevant use case. The use case has provided evidence that the proposed concept can be used for training vision-based recognition of parts’ orientation using simulation of DT models, which in turn can be used for adaptively controlling the production process.