New data-driven predictive modelling methods for data scarcity scenarios in smart manufacturing

Data-driven smart manufacturing has demonstrated tremendous potential across the entire manufacturing lifecycle, and initiated and enriched a series of new paradigms, such as digital twin, Industry 4.0, cloud manufacturing and IoT-enabled manufacturing. As a fundamental topic of smart manufacturing, manufacturing predictive modelling (MPM) aims to construct high-fidelity predictive representations of the concerned properties of products, processes or manufacturing systems, for further process optimisation for decision-making.Various machine-learning methods have been developed and successfully applied to address different manufacturing predictive modelling problems. However, the superior performance and effectiveness of data-driven predictive modelling methods heavily depend on the availability of a substantial amount of labelled data. Unfortunately, labelling manufacturing data, both computationally and experimentally, is often a costly and time-consuming task. Therefore, establishing data-driven predictive models with scarce and insufficient labelled data is an inevitable challenge for the development of smart manufacturing.Generally, the performance limit for data driven predictive models is constrained by the information contained within the labelled dataset, particularly for data scarcity scenarios. This leads to the research gap, the insufficient modelling information. Therefore, it is necessary to actively leverage the data generation process, while simultaneously compensating for the insufficient modelling information by incorporating other available sources of information.To deal above-mentioned research gap for data driven predictive modelling under data scarcity, this thesis proposed a systematic framework including the active sampling of direct labelled data, knowledge transfer from the auxiliary data and data-physics combination. (a) Firstly, an aggregation-value-based sampling (AV4Sam) method was proposed based on the Game theory for sampling the most promising labelled data. Experiments on several manufacturing cases demonstrated that the proposed method could generate optimal dataset actively compared with state-of-arts methods As a result, the data generation process can be proactively harnessed to facilitate the subsequent data modelling process. (b) A new transfer learning approach, structured Conditional Distribution Adaptation (CDA), was proposed to facilitate the knowledge transfer from auxiliary data to the target task, thus improving the performance of the target model under data scarcity situations. (c) To further leverage the wide-existed physics prior knowledge of the manufacturing process, a physics-guided Low-dimensional Neural Operator (LNO) was developed, which incorporates physics priors into the neural network structure to enhance the learning capabilities in predicting the high-dimensional part properties.The proposed methods of data sampling, transfer learning and data-physics combination were validated in various manufacturing cases including composite curing, multi-sensor measurement and tool wear prediction. In addition, a data-driven curing deformation prediction system was developed for composite manufacturing based on the proposed methods. The effectiveness of the system was validated in the curing deformation prediction of a complex composite workpiece. Experimental results show that the developed system could provide accurate deformation prediction results while significantly reducing the required training data by 90% compared to the existing method.In summary, this thesis leverages manufacturing data generating process to enhance the data-driven predictive modelling under data scarcity scenarios. The proposed methods not only contribute to the engineering-oriented machine learning, but also offer valuable insights for the development of smart manufacturing.