[{"data":1,"prerenderedAt":226},["ShallowReactive",2],{"content-query-BlkQ4eOnEs":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"body":13,"_type":220,"_id":221,"_source":222,"_file":223,"_stem":224,"_extension":225},"/news/zh/3813","zh",false,"","东南大学魏秀参：基于智能感知的工业AI质检算法与应用实践","形成“发现问题——分析问题——解决问题” 的思维闭环","2025-08-14","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/08/22/13cf9abd792e4463b3455fc7c277cd29.png","news",{"type":14,"children":15,"toc":217},"root",[16,24,33,39,44,49,57,62,67,72,77,82,90,95,100,105,110,115,120,125,134,139,144,149,154,161,166,171,176,181,186,194,199,207,212],{"type":17,"tag":18,"props":19,"children":21},"element","h1",{"id":20},"东南大学魏秀参基于智能感知的工业ai质检算法与应用实践",[22],{"type":23,"value":8},"text",{"type":17,"tag":18,"props":25,"children":27},{"id":26},"项目实施背景",[28],{"type":17,"tag":29,"props":30,"children":31},"strong",{},[32],{"type":23,"value":26},{"type":17,"tag":34,"props":35,"children":36},"p",{},[37],{"type":23,"value":38},"近年来，随着智能制造和工业自动化的加速发展，产品质量检测已成为制造业高质量发展的关键，企业对质检效率、准确率及数据可追溯性提出了更高要求。随着人工智能技术的快速进步，智能感知技术构建了覆盖产品外观、结构和性能等多方面的高精度感知体系，为工业质检系统提供了更精准高效的检测手段，尤其是在视觉质检领域，智能感知技术极大提升了检测速度与一致性，显著降低了人力成本和误检风险，已成为电子、汽车等高端制造行业不可或缺的重要工具。",{"type":17,"tag":34,"props":40,"children":41},{},[42],{"type":23,"value":43},"然而，当前工业AI质检在实际应用过程中仍面临诸多挑战：一方面，在实际工业生产中，异常样本往往稀缺且分布复杂，数据采集和标注成本高昂，导致模型泛化能力有限，难以有效应对多变的生产场景；另一方面，虽然基于正样本的异常检测方法已应用，但该方法仍存在类别边界模糊、缺乏参照样本以及对未知异常识别能力不足等问题，直接影响其实用性和可靠性。",{"type":17,"tag":34,"props":45,"children":46},{},[47],{"type":23,"value":48},"针对正样本的异常检测理论与应用创新，本项目提出新的方法体系，通过理论创新和技术攻关，推动质检的智能化、自动化和规模化发展；并将项目成果融入课堂教学，打造真实的学习情境，形成学生 “发现问题——分析问题——解决问题” 的思维闭环，提升学生科研探索的精神以及解决实际问题的能力。",{"type":17,"tag":18,"props":50,"children":52},{"id":51},"改革思路及举措",[53],{"type":17,"tag":29,"props":54,"children":55},{},[56],{"type":23,"value":51},{"type":17,"tag":34,"props":58,"children":59},{},[60],{"type":23,"value":61},"本项目以基于智能感知的工业AI质检算法为核心，围绕工业现场实际需求，聚焦多层次信息融合、深度特征挖掘、异常检测等关键技术环节，引入元学习（Meta-Learning）思想，在极小样本条件下，通过智能算法实现对单类目标的精准高效识别（One-Shot One-Class Classification， OSOC），有效突破现有AI质检对大量异常样本的依赖。此外，项目构建了基于智能感知的全流程AI质检范式，实现理论创新与工程落地的统一。",{"type":17,"tag":34,"props":63,"children":64},{},[65],{"type":23,"value":66},"1、理论创新：极小样本单类识别的元学习范式",{"type":17,"tag":34,"props":68,"children":69},{},[70],{"type":23,"value":71},"针对数据稀缺的工业环境，借助元学习框架，通过大量模拟任务的训练，使模型具备“由单一样本自动推断类别边界”的能力，即在每一个元训练任务中，系统随机抽取一个目标类别，并仅以该类别的一个样本作为“示例”，再从该类别及其“近邻类别”中采样组成测试集，模型基于该示例自动生成判别分类器及参考分数，实现同类样本与异类样本的准确区分。",{"type":17,"tag":34,"props":73,"children":74},{},[75],{"type":23,"value":76},"2、方法创新：层次多粒度采样与单类判别",{"type":17,"tag":34,"props":78,"children":79},{},[80],{"type":23,"value":81},"针对工业质检中常见的细粒度类别混淆，在任务构建环节提出层次多粒度采样方法：以辅助数据集的类别层级结构（如WordNet语义树）为依据，模拟现实中的“近邻类别混淆”，针对性地从目标类别的近邻类别中采样负样本，显著提升了模型在细粒度识别中的适应性和判别精度。",{"type":17,"tag":34,"props":83,"children":84},{},[85],{"type":17,"tag":86,"props":87,"children":89},"img",{"alt":7,"src":88},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/08/22/caada8b683e54a30b6de0adaae7290dd.png",[],{"type":17,"tag":34,"props":91,"children":92},{},[93],{"type":23,"value":94},"图1 元学习框架下单类分类判别机理",{"type":17,"tag":34,"props":96,"children":97},{},[98],{"type":23,"value":99},"在模型结构设计方面，通过非线性多层感知机完成由示例样本特征到分类器权重的映射。同时，基于该示例自动生成参考分数，作为后续类别判别的阈值，在实际判别过程中，将测试样本特征与分类器权重做内积，得到属于目标类别的得分，并与参考分数比较，得分高于阈值即判为目标类别，反之为非目标类别。",{"type":17,"tag":34,"props":101,"children":102},{},[103],{"type":23,"value":104},"3、优化策略：多目标自适应阈值调节",{"type":17,"tag":34,"props":106,"children":107},{},[108],{"type":23,"value":109},"为确保模型能在极小样本下捕捉类别本质特征，设计了多目标优化损失函数。一方面，采用基于参考分数的排序损失，使目标类别样本得分高于阈值，非目标类别样本得分低于阈值，从而有效拉开正负样本间决策边界；另一方面，引入示例置信度约束，推动分类器在示例样本上输出最大置信得分，提升对类内差异的自适应能力。",{"type":17,"tag":34,"props":111,"children":112},{},[113],{"type":23,"value":114},"针对阈值过高可能导致召回率下降的问题，项目引入了“自适应冗余变量”机制：即在训练过程中，模型可根据示例样本特性动态调整参考分数，实现判别能力与召回率的自动平衡，大幅提升了模型对类别多样性和实际环境变化的鲁棒性。",{"type":17,"tag":34,"props":116,"children":117},{},[118],{"type":23,"value":119},"4、工程落地：全流程端到端优化",{"type":17,"tag":34,"props":121,"children":122},{},[123],{"type":23,"value":124},"采用全流程端到端优化思路，充分发挥深度学习在特征表达与非线性映射方面的优势，实验结果显示，所提出的方法在多个OSOC基准数据集上均明显优于传统阈值法、OC-SVM（One-Class Support Vector Machine）、简单映射法及主流CLEAR方法，尤其在细粒度识别与异常检测场景下表现突出。同时，模型设计兼顾了工程实现的简洁性和高效性，便于在实际工业质检系统中集成和部署。",{"type":17,"tag":18,"props":126,"children":128},{"id":127},"项目成果创新点及成效",[129],{"type":17,"tag":29,"props":130,"children":131},{},[132],{"type":23,"value":133},"项目成果、创新点及成效",{"type":17,"tag":34,"props":135,"children":136},{},[137],{"type":23,"value":138},"1、理论与方法创新",{"type":17,"tag":34,"props":140,"children":141},{},[142],{"type":23,"value":143},"针对工业AI质检场景中异常样本难以收集、单类样本建模准确性不足等核心难题，本项目以智能感知技术为基础，创新性地将元学习与单类分类相结合，提出并实现了基于正样本的类别边界生成、元排序学习等核心算法，在无或极少量异常样本条件下，精准捕捉产品的异常特征，显著提升检测性能。同时，相关技术在细粒度异常检测、罕见类别识别等任务上展现出优越的适应性和泛化能力。",{"type":17,"tag":34,"props":145,"children":146},{},[147],{"type":23,"value":148},"2、产学融合与应用",{"type":17,"tag":34,"props":150,"children":151},{},[152],{"type":23,"value":153},"充分依托昇思MindSpore框架，紧密将企业实际需求与高校科研创新结合，在智能感知技术与算法研究、工程实践等方面做了深度融合。围绕工业AI质检中的实际应用问题，项目团队以基于类别元型的单类分类算法为核心，完成了从智能感知、数据预处理、算法设计、平台实现到实际应用的全链条攻关。基于昇思MindSpore实现了高效的单类分类算法，并在视觉数据集MVTecAD（MVTec Anomaly Detection）上进行了系统性评测，验证了算法的先进性和可用性。",{"type":17,"tag":34,"props":155,"children":156},{},[157],{"type":17,"tag":86,"props":158,"children":160},{"alt":7,"src":159},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/08/22/9bc530befc714c1ea02c2b6307208314.png",[],{"type":17,"tag":34,"props":162,"children":163},{},[164],{"type":23,"value":165},"图2 基于正样本的异常检测算法(FastFlow)端到端方案",{"type":17,"tag":34,"props":167,"children":168},{},[169],{"type":23,"value":170},"同时，项目团队基于昇思MindSpore和智能感知系统实现了支持基于正样本异常检测的端到端检测方案FastFlow。该方法通过FastFlow模块对图像特征分布进行重建和建模，将正常产品的特征映射到分布中心，异常特征则远离分布中心，实现了高效、精准的异常检测。相关算法框架可在端、边、云等多种工业应用环境下灵活部署，极大提升了工业质检系统的智能化、自动化水平。项目成果已在部分工业企业试点应用，取得了良好的经济效益和社会反响。",{"type":17,"tag":34,"props":172,"children":173},{},[174],{"type":23,"value":175},"3、人才培养与教育改革",{"type":17,"tag":34,"props":177,"children":178},{},[179],{"type":23,"value":180},"依托项目建设，团队积极推进“产学研融合”创新人才培养模式。项目团队成员涵盖高校教师、博士研究生及企业研发人员，形成了多元协同的创新团队；项目高度重视对学生科研创新能力与实践能力的培养，鼓励并支持学生全过程、深层次参与课题研究，包括实际工程问题分析、算法设计与优化、系统实现、实验验证等各个环节，在项目实施过程中，学生不仅积极参与高水平论文的撰写与国际学术交流，还在导师和企业工程师的指导下，联合申报发明专利和软件著作权，培养了从基础理论到技术转化的全链条创新能力。团队定期组织学术沙龙、技术竞赛和企业实习，促进学生与企业研发团队的深度交流与协作，推动产学合作的有效落地与循环提升，实现科研创新、技术转化和人才培养的统一。",{"type":17,"tag":34,"props":182,"children":183},{},[184],{"type":23,"value":185},"同时，团队将项目核心方法集成到课程建设与实践教学中，开发了细粒度图像分析等系列课程，推动元学习、小样本智能识别和智能感知领域的理论与实践一体化教学，提升了学生在智能感知与工业AI质检领域的创新能力和工程实践能力，为行业输送了高素质复合型人才。",{"type":17,"tag":18,"props":187,"children":189},{"id":188},"应用及推广情况",[190],{"type":17,"tag":29,"props":191,"children":192},{},[193],{"type":23,"value":188},{"type":17,"tag":34,"props":195,"children":196},{},[197],{"type":23,"value":198},"在重大工程领域，项目成果已实现应用突破：团队开发了环焊缝缺陷识别技术，并已在中国海油的海底管道检测工作中试点部署。该技术针对海底管道环焊缝缺陷类型多、人工检测难度大等实际问题，应用了基于正样本的AI识别模型，实现对高危管道环焊缝缺陷的高精度定位，显著提升了检测效率与准确率，为及时识别和处置潜在安全隐患提供了坚实保障。",{"type":17,"tag":18,"props":200,"children":202},{"id":201},"经验总结",[203],{"type":17,"tag":29,"props":204,"children":205},{},[206],{"type":23,"value":201},{"type":17,"tag":34,"props":208,"children":209},{},[210],{"type":23,"value":211},"本项目以智能感知为核心，实现了多源传感与AI算法的深度融合，显著提升了工业质检的智能化和自动化水平，形成“技术研发——产业应用——人才培养”的良性循环。未来，团队将持续优化智能感知技术，拓展多样化工业场景应用，推动工业AI质检技术不断向更高水平发展。",{"type":17,"tag":34,"props":213,"children":214},{},[215],{"type":23,"value":216},"案例内容由项目负责人提供",{"title":7,"searchDepth":218,"depth":218,"links":219},4,[],"markdown","content:news:zh:3813.md","content","news/zh/3813.md","news/zh/3813","md",1776506089914]