面向可解释工业设备异常诊断的多模态大模型框架
KEHR-IAD: A Multimodal Large Language Model Framework for Interpretable Industrial Equipment Health Diagnosis
娄骁扬 1章洋1
作者信息
- 1. 北京邮电大学计算机学院，北京 100876
- 折叠
摘要
针对当前多模态大模型在工业异常检测任务中面临的领域知识匮乏及细微瑕疵感知能力不足的问题,提出一种基于知识增强混合推理的检测方法。该方法首先构建多层次特征融合机制,将深度视觉网络的细粒度特征与大模型生成的可解释语义特征相结合,以增强对微小缺陷的表征能力。其次,设计融合检索增强生成与思维链的推理范式,通过引入领域知识库并规范推理路径,实现对异常的精准诊断。在主流工业异常检测数据集上的全样本与少样本设置实验结果表明,该方法在检测精度与泛化性能上显著优于现有主流方法。研究证实,所提方法能有效抑制模型幻觉并提升诊断的可解释性,为大模型在工业质检场景的落地应用提供了新思路。
Abstract
This paper addresses the challenges faced by Multimodal Large Language Models in Industrial Anomaly Detection tasks, specifically the lack of domain-specific knowledge and insufficient perception of subtle defects. A knowledge-enhanced hybrid reasoning detection method is proposed to tackle these issues. The method first constructs a multi-level feature fusion mechanism, combining fine-grained features extracted from deep visual networks with interpretable semantic features generated by large models, thereby enhancing the representation capability for subtle defects. Secondly, a reasoning paradigm integrating Retrieval-Augmented Generation and Chain-of-Thought is designed. This paradigm achieves accurate diagnosis of anomalies by incorporating a domain knowledge base and standardizing reasoning paths. Experimental results on mainstream industrial anomaly detection datasets, evaluated under both full-shot and few-shot settings, demonstrate that the proposed method significantly outperforms existing state-of-the-art approaches in detection accuracy and generalization performance. This study confirms that the proposed method effectively mitigates model hallucination and enhances diagnostic interpretability, offering a novel avenue for the practical application of large models in industrial quality inspection scenarios.关键词
人工智能/工业异常检测/多模态大模型Key words
AI/Industrial Anomaly Detection/Multimodal Large Language Models引用本文复制引用
娄骁扬,章洋.面向可解释工业设备异常诊断的多模态大模型框架[EB/OL].(2026-03-12)[2026-03-15].http://www.paper.edu.cn/releasepaper/content/202603-112.学科分类
自动化技术、自动化技术设备/计算技术、计算机技术
评论