Cardiff University | Prifysgol Caerdydd ORCA
Online Research @ Cardiff 
WelshClear Cookie - decide language by browser settings

Knowledge distillation for energy consumption prediction in additive manufacturing

Li, Yixin, Hu, Fu, Ryan, Michael ORCID:, Wang, Ray and Liu, Ying ORCID: 2022. Knowledge distillation for energy consumption prediction in additive manufacturing. Presented at: 14th IFAC Workshop on Intelligent Manufacturing Systems (IMS 2022), Tel-Aviv, Israel, 28-30 March 2022. IFAC-PapersOnLine. IFAC-PapersOnLine. , vol.55(2) Elsevier, pp. 390-395. 10.1016/j.ifacol.2022.04.225

[thumbnail of 1-s2.0-S2405896322002269-main.pdf]
PDF - Published Version
Available under License Creative Commons Attribution Non-commercial No Derivatives.

Download (676kB) | Preview


Owing to the advances of data sensing and collecting technologies, more production data of additive manufacturing (AM) systems is available and advanced data analytics techniques are increasingly employed for improving energy management. Current supervised learning-based analytical methods, however, typically require extracting and learning valuable information from a significant amount of data during training. It is difficult to make a trade-off between latency and computing resources to implement the analytical models. As such, this paper developed a method utilizing the knowledge distillation (KD) technique for predicting AM energy consumption based on product geometry information to reduce computational burdens while simultaneously retaining model performance. Through a teacher-student architecture, layer-by-layer images of products and energy consumption datasets are used to train a teacher model from which the knowledge is extracted and used to build a student model to predict energy consumption. A case study was conducted to demonstrate the feasibility and effectiveness of the proposed approach using real-world data from a selective laser sintering (SLS) system. Comparisons between distilled and independently trained student models were made in terms of the root mean square error (RMSE) and training time. The distilled student model performed better (14.3947KWh/kg) and required a shorter training time (34s) than the complex teacher model.

Item Type: Conference or Workshop Item (Paper)
Date Type: Publication
Status: Published
Schools: Engineering
Publisher: Elsevier
ISSN: 2405-8963
Date of First Compliant Deposit: 16 December 2021
Last Modified: 29 Nov 2022 10:22

Actions (repository staff only)

Edit Item Edit Item


Downloads per month over past year

View more statistics