Design of an industrial quality inspection system based on a large model for multimodal data fusion and anomaly detection

Abstract

To address the issues of inefficiency and poor adaptability caused by traditional industrial visual inspection—which relies on manual interpretation and fixed rules—this paper designs and implements an industrial quality inspection system that integrates multi-modal data with the anomaly detection capabilities of large vision models. The system fuses RGB and other multi-modal data and employs unsupervised anomaly detection with RAM, Grounding, and AdaptCLIP to achieve few-shot defect localization and quantification.By incorporating AdaptCLIP and Grounding as the core reasoning engine, the system dynamically generates strategies through comprehensive reasoning based on product labels, system parameters, and image data. An eight-step standardized quality inspection process and a five-level response mechanism based on anomaly area and location enable progressive decision-making from product release to rework and repair. The system’s effectiveness has been validated in the inspection of automotive logos and wine bottle labels. The integrated "decision-detection-recording" solution offers a practical technical pathway for quality inspection in smart manufacturing.