Cross-modal and sparse activation-based knowledge graph construction and reasoning for process planning

Abstract

Process planning is a key enabler of manufacturing automation and intelligence, and it is especially critical to the quality and efficiency of complex-product manufacturing. However, due to departmental isolation, fragmentation of heterogeneous process knowledge, interference from task-irrelevant nodes, and insufficient reasoning focus, existing process knowledge models often suffer from incomplete representation, inefficient inference, and imprecise recommendation. To address these limitations, this paper proposes a cross-modal and sparse activation-based knowledge graph construction and reasoning method for process planning. First, process data and knowledge from multiple departments and heterogeneous modalities are unified into a triple-based representation for constructing a process knowledge graph. Second, a sparse activation mechanism-based minimum-cost reasoning model is developed for static process knowledge inference, enabling the extraction of a highly relevant subgraph and the computation of the minimum-cost process route from a start node to an end node. Third, a large language model (LLM)-based interactive question-answering recommendation framework is proposed to generate a single high-scoring answer from a dynamically updated process knowledge graph in real time. In addition, the proposed framework is closely related to the emerging paradigm of physics-informed machine learning, because process planning decisions are inherently constrained by physical mechanisms, process rules, and manufacturing knowledge, which can be explicitly encoded and activated through the knowledge graph. Finally, an actual production scenario-based case study and extensive experiments are conducted to validate the effectiveness and superiority of the proposed methodology.