Active transfer learning for data-driven manufacturing process modelling

Manufacturing process modelling (MPM) aims to construct high-fidelity digital predictive models of the concerned properties of products, processes or manufacturing systems for the further optimisation and improvement of manufacturing activities. Data-driven modelling methods, including machine learning and deep learning, have drawn immense attention to MPM problems because of their powerful representative ability. However, the labelled data of concerning properties in the manufacturing process is often insufficient and sparse because of the expensive and time-consuming experiments or simulations. The scarcity of labelled data hinders the further development of data-driven models in MPM problems. This paper proposes an active transfer learning framework by integrating active generation of labelled data and the processing of relevant data to reduce the requirements of labelled data. Firstly, the initial active labelling module introduces the generation of a more representative and informative labelled dataset rather than a randomly generated one. Then, the transfer learning model can extract the general information from the relevant data to address the information scarcity for the target task. Besides, the iterative active labelling module can determine to query promising new labelled data according to the performance of the current model. The effectiveness of the proposed framework is verified in a tool wear prediction case. The experimental outcomes demonstrate that the three modules of the framework can reduce the labelled data requirements and enhance the performance of the data-drive model under limited labelled data.