Stage-aware industrial defect understanding via multi-agent collaboration

Defect detection is a fundamental task in industrial manufacturing, traditionally focused on locating defect regions and classifying their types. However, existing methods often overlook deeper semantic understanding, especially the comprehensive information of defects such as their causes, impacts, and possible solutions, which may vary significantly across different supply chain stages. To bridge this gap, we propose a novel task called stage-aware industrial defect understanding, which encourages models to learn to understand the comprehensive information of defects across different supply chain stages. To support this task, we construct a new dataset that integrates images of defects and their corresponding supply chain stages, and adopts a carefully designed visual question answering (VQA) format covering three levels of understanding, i.e., defect type recognition, stage-context reasoning, and comprehensive information analysis. Considering the multi-step structure of our task, which spans from basic defect recognition to comprehensive information understanding across different supply chain stages, we propose a multi-agent framework named stage-aware VQA (SA-VQA). This framework consists of four collaborative agents, i.e., defect discerner, stage concluder, stage predictor, and summarize. Each agent is powered by large vision–language models (LVLM) and interacts via a shared information pool, ensuring coherent reasoning and interpretable outputs. Experimental results demonstrate that SA-VQA significantly outperforms strong baselines across all levels of questions, especially in complex reasoning scenarios involving stage-aware semantics. This work lays a solid foundation for advancing defect analysis from basic perception to semantic-level understanding in real-world industrial applications.