数据与计算发展前沿 ›› 2025, Vol. 7 ›› Issue (6): 179-190.

CSTR: 32002.14.jfdc.CN10-1649/TP.2025.06.017

doi: 10.11871/jfdc.issn.2096-742X.2025.06.017

• 技术与应用 • 上一篇    

面向智慧供应链业务融合的轻量化多模态物料识别与追溯技术研究

阮敏1(),王琳2,张爱霞3,*(),王望龙3   

  1. 1.国能数智科技开发（北京）有限公司，北京 100010
    2.国能置业有限公司，北京 100010
    3.北京中科院软件中心有限公司，北京 100190
  • 收稿日期:2025-06-26 出版日期:2025-12-20 发布日期:2025-12-17
  • 通讯作者: 张爱霞
  • 作者简介:阮敏，国能数智科技开发（北京）有限公司，高级工程师，主要研究与工作方向为ERP、供应链管理应用。
    本文承担工作为：研究方案设计，提出轻量化多模态蒸馏与可信溯源融合的创新技术架构，确定技术可行性与实施路径。组织完成图像识别算法核心开发、实验设计，撰写论文技术架构章节核心内容。
    RUAN Min, senior engineer of CHN energy digital intelligence technology development (Beijing) Co., Ltd. Her research interests include ERP and supply chain management application.
    In this paper, she is responsible for the design of the research framework, proposing an innovative technical framework integrating lightweight multimodal distillation and blockchain traceability, and verifying the technical feasibility as well as the implementation pathway. She also organizes and completes the core development of the image recognition algorithms and the experimental design, and writes the main content of the technical framework section of the paper.
    E-mail: 11688403@ceic.com|张爱霞，北京中科院软件中心有限公司，工程师，主要研究与工作方向为人工智能应用、工业仿真优化与计算等。
    本文承担工作为：完成图像识别算法核心开发、实验设计、实现及结果分析，撰写论文实验效果分析章节内容。
    ZHANG AiXia, engineer of Software Engineering Center Chinese Academy of Sciences. Her research interests include artificial intelligence applications, industrial simulation optimization and computing.
    In this paper, she is responsible for completing the core development, experimental design, implementation, and result analysis of the image recognition algorithms, and for writing the experimental results and analysis section of the paper.
    E-mail: axzhang@sec.ac.cn
  • 基金资助:
    国家重点研发计划(2022YFD00605)

An Study on Integration of Lightweight Multimodal Distillation and Trusted Traceability for Ingredient Recognition in Smart Supply Chain

RUAN Min1(),WANG Lin2,ZHANG Aixia3,*(),WANG Wanglong3   

  1. 1. Digital and Intelligent Technology Development (Beijing) Co., Ltd of CHINA ENERGY GROUP, Beijing 100010, China
    2. Real Estate Investment Co., Ltd of CHINA ENERGY GROUP, Beijing 100010, China
    3. Beijing Software Engineering Center Co., Ltd, Chinese Academy of Sciences, Beijing 100190, China
  • Received:2025-06-26 Online:2025-12-20 Published:2025-12-17
  • Contact: ZHANG Aixia

RichHTML

1

PDF

0

摘要：

【目的】针对后勤食材供应链出入库场景下物料识别存在的细粒度分类难、溯源可信度低等问题，本文提出轻量化多模态蒸馏与区块链溯源融合的创新技术架构，内置业务约束辅助提升食材物料的细粒度分类识别，通过物料图像-采购订单-供应商合同等数据的可信溯源来确保相关业务的合规性，为智慧后勤食材供应链数字化转型提供参考。【方法】区别于传统的多模态识别，该架构打破局限，构建跨模态注意力对齐的动态蒸馏框架，将采购台账语义约束嵌入YOLOv7-tiny改进模型，通过视觉-称重-业务规则三重特征融合，实现部分遮挡场景下的食材物料识别。进一步改进双链式区块链技术框架，通过联邦学习优化智能合约，实现食材称重图像的快速溯源验证，形成“精准识别-动态优化-可信溯源”的技术闭环。【结果】有效提升了食材物料识别精度，有效确保了数据防篡改和系统稳定性，降低采购偏差，满足食材安全品控要求。【局限】对部署环境有一定要求，在后续研究中可持续实验并提出可行性方案。【结论】为智慧后勤供应链区块链、人工智能应用提供了具有可复制性的新范式。有望在国央企中进行推广，有效减少年损耗成本，助力后勤采购实现精准可信，高品高效。

关键词: 智慧供应链, AI智能云秤, 轻量化, 知识蒸馏, 区块链溯源

Abstract:

[Purpose] Addressing challenges such as difficult fine-grained classification and low traceability credibility in material identification from the inventory scenarios of logistics food supply chain, this paper proposes an innovative technical framework integrating lightweight multimodal distillation and blockchain traceability. Built-in business constraints assist in improving the fine-grained classification and recognition of food materials, ensuring compliance of related businesses through reliable traceability of data such as material images, purchase orders, and supplier contracts, providing reference for the digital transformation of the smart logistics food supply chain. [Methods] Different from traditional multimodal recognition, this technical framework breaks through limitations and constructs a dynamic distillation framework with cross-modal attention alignment. The semantic constraints of the procurement ledger are embedded in the YOLOv7 tiny improved model. Through the triple feature fusion of visual, weighing, and business rules, food material recognition in partially occluded scenes is achieved. Further improvement of the dual chain blockchain technology framework and optimization of the smart contracts through federated learning achieve rapid traceability verification of food weighing images, forming a technical loop of "precise recognition, dynamic optimization, and trusted traceability". [Results] This work effectively improves the accuracy of food ingredient identification, ensures data tamper resistance and system stability, reduces procurement deviation, and meets the requirements of food safety and quality control. [Limitations] There are certain specific requirements for the deployment environment, and further experiments will be conducted in follow-on research to propose feasible solutions. [Conclusions] This work provides a replicable new paradigm for smart logistics supply chain using blockchain and artificial intelligence applications. It is expected to be promoted in state-owned enterprises, effectively reducing the cost of yearly loss and improving logistics procurement precision, reliability, quality and efficiency.

Key words: smart supply chain, smart AI cloud scale, lightweighting, knowledge distillation, blockchain traceability