Edge Summaries and Multimodal Prompt Feedback Advance Factory Defect Inspection with Vision Language Models Systems

Abstract

The rapid evolution of modern manufacturing paradigms demands highly efficient, scalable, and adaptive quality control mechanisms. Traditional centralized cloud-based defect inspection systems often suffer from significant latency, bandwidth constraints, and an inability to rapidly adapt to newly emerging defect typologies on the factory floor. To address these critical limitations, this paper proposes a novel framework for factory defect inspection that synergizes edge computing architectures with advanced large vision-language models through a mechanism termed Edge Summaries and Multimodal Prompt Feedback. In this paradigm, raw high-fidelity sensory and visual data captured at the manufacturing line are compressed and semantically distilled into lightweight edge summaries by localized computing nodes. These summaries are then transmitted to a centralized, multimodal analytical engine, which evaluates the structural and semantic integrity of the manufactured components. Based on this analysis, the central model generates context-aware multimodal prompt feedback, which is transmitted back to the edge nodes to dynamically adjust detection thresholds, refine feature extraction parameters, and adapt to environmental variations such as lighting changes or novel anomaly presentations. Through extensive theoretical analysis and architectural design, this paper details the operational workflow, data synchronization protocols, and adaptation mechanisms of the proposed system. The findings suggest that employing edge summaries dramatically reduces bandwidth consumption while the prompt feedback loop maintains the high analytical accuracy typically associated with massive centralized models. This research provides a foundational blueprint for deploying resilient, intelligent, and highly responsive quality assurance systems in Industry 4.0 environments.