CausalKGPT: Industrial structure causal knowledge-enhanced large language model for cause analysis of quality problems in aerospace product manufacturing

The whole cycle for manufacturing aerospace thin-walled shells is a lengthy and sophisticated process. A large amount of quality-related data exists within and between processes, involving many types of quality defects and influencing factors. However, there are ambiguous causal associations among quality-related data affecting the shape-properties of the shell. Also, the coupling of long processes and multiple factors makes it hard to analyze the main factors that affect the quality defects in shell manufacturing. In this paper, taking into account the advantages of causal Scientology and the large language model (LLM), we propose an industrial structure causal knowledge-enhanced large language model for the cause analysis of quality defects in aerospace product manufacturing. To reinforce the causal associations among quality-related data deriving from manufacturing documents (product defect survey sheets, quality inspection, and maintenance reports), a structure causal graph-based sum-product network (SCG-SPN) model is designed to model machining quality-related knowledge and eliminate pseudo-association confounding factors by doing an intervention. Thus, a causal quality-related knowledge graph (CQKG) with high-quality causal associations is constructed. With this, to provide a trustworthy guarantee in responding to quality problem solving, we construct a quality-related prompt dataset with multi-round conversations based on CQKG. Then, a novel P-tuning that adapts to utilize external CQKG instructions is designed to fine-tune an open-source ChatGLM base model. Based on this, a causal knowledge graph-augmented LLM, named CausalKGPT, is developed to enable reasoning and responding to quality defects in both Chinese and English. It uses natural text descriptions related to quality defects as input and takes a quality-related causal knowledge graph as an additional corpus. Finally, the case study shows that the CausalKGPT performs with more expertise and reliability in responding to quality question solving of aerospace shell manufacturing than the classic commercial models like ChatGPT and GPT4. The results indicate that the proposed method may provide a trustworthy guide in assisting workers to analyze quality defects in aerospace products.