Task incremental learning-driven Digital-Twin predictive modeling for customized metal forming product manufacturing process

• Published in: Robotics and computer-integrated manufacturing Vol. 85; p. 102647
• Main Authors: Li, Jie, Wang, Zili, Zhang, Shuyou, Lin, Yaochen, Jiang, Lanfang, Tan, Jianrong
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.02.2024
• Subjects: custom product manufacturing; Digital-Twin; metal forming; predictive modeling; task incremental learning; Temporal Fusion Transformer; Digital-Twin; task incremental learning; Temporal Fusion Transformer; metal forming; predictive modeling; custom product manufacturing
• ISSN: 0736-5845; 1879-2537
• DOI: 10.1016/j.rcim.2023.102647

•Digital-Twin model for custom metal forming process quality prediction was developed.•Comprehensive information model was created for integrating multi-source information.•Temporal fusion transformer was proven to be able to mechanical process prediction.•Task incremental learning was adopted for rapid and scalable predictive modeling.•Digital-Twin method accurately predicted ovality in the tube bending process. Customized metal forming products entail personalized requirements in terms of dimensions, materials, and other specifications, while the processing conditions involved are subject to dynamic changes. Digital-Twin (DT) predictive models have become essential tools for optimizing the complex manufacturing process. However, the traditional approach exhibits limitations in handling dynamic data, capturing complex nonlinear relationships, and leveraging multi-source information. Additionally, retraining predictive models for novel tasks with unique operating conditions in specific scenarios can lead to substantial time and resource inefficiencies. Therefore, a task incremental learning-based approach for DT predictive modeling is proposed in this paper. Firstly, a DT framework and a comprehensive information model are established for real-time monitoring and integration of multi-source information. Moreover, the pre-trained Temporal Fusion Transformer model is utilized to capture valuable knowledge from historical tasks. Subsequently, task incremental learning is adopted to fine-tune the model using new task data, thereby enhancing adaptability and enabling rapid and scalable modeling. Finally, the effectiveness of the proposed method is validated on a customized metal tube bending forming platform, demonstrating accurate prediction of tube cross-section deformation.