Multidomain Graph Meta-Learning Network for Few-Shot Prediction in Industrial Processes

In process industries, the scarcity of data highlights the necessity of few-shot learning for accurate industrial predictions. Recently, model-agnostic meta-learning (MAML) has become an effective solution to this challenge. However, the existing MAML in industrial prediction still faces two key issues: 1) training tasks originate from multiple domains and 2) the intervariable coupling relationships are not fully exploited to their potential in cases with limited information. Therefore, we propose a multidomain graph meta-learning network (MDGML) to enable knowledge transfer and enhance information utilization through domain generalization and graph-based meta-learning. First, an empirical domain adaptation module is developed to map each source domain to a shared domain, ensuring consistent distributions across learning tasks. Next, robust hyperbolic graphs are devised to learn the spatial coupling relationships among process variables and offer additional insights through a graph-based perspective. Then, domain-specific prototypes are computed and designed to modulate the prediction results, thereby enhancing domain-dependent prediction accuracy. Furthermore, we also propose task-adaptive optimization for MAML inner loops under the regulation of the node information density (NID) in graphs. It can accommodate differences between learning tasks. Finally, the experimental results from four real-world blast furnace datasets validate the effectiveness of the proposed method.