基于半监督学习双模型结构的注塑产品异常检测  

Abnormal Detection of Injection Molding Products Based on Semi-supervised Learning Dual-model Structure

作　　者：陈昱 项薇[1,2] 林文文 龚川[1] 张怀志 虞任豪 CHEN Yu;XIANG Wei;LIN Wenwen;GONG Chuan;ZHANG Huaizhi;YU Renhao(School of Mechanical Engineering and Mechanics,Ningbo University,Ningbo,Zhejiang,315211;Zhejiang Provincial Key Laboratory of Part Rolling Technology,Ningbo,Zhejiang,315211)

机构地区：[1]宁波大学机械工程与力学学院,宁波315211 [2]浙江省零件轧制成形技术研究重点实验室,宁波315211

出　　处：《中国机械工程》2025年第3期576-583,共8页China Mechanical Engineering

摘　　要：质量数据分布的不平衡及分类边界的模糊性限制了传统分类器的性能,阻碍了企业智能生产决策的高效实施。为此,提出了一种基于双模型结构的深度生成模型异常检测方法,根据尺寸数据分布将合格产品等级进行二分类,即优秀及次优,分别用于训练两个深度生成模型,考虑数据分布特点设计加权集成,基于计算的异常分数对产品进行合格性判定。以变分自编码器(VAE)、Wasserstein生成对抗网络(WGAN)为子模型开发了两个双模型结构,测试结果显示,相较于单模型结构,基于双模型的VAE和WGAN在测试集上的分类准确率分别提高了4.5%和6%。The imbalance in quality data distribution and the ambiguity of classification boundaries were identified as the factors limiting the performance of traditional classifiers,which hindered the efficient implementation of intelligent production decision-making in enterprises.Therefore,a dual-model-based deep generative model anomaly detection method was proposed.Qualified products were classified into two levels,excellent and sub-optimal,based on size data distribution,and two deep generative models were trained separately.A weighted ensemble strategy was designed considering characteristics of data distribution,and anomaly scores were calculated to determine product qualification.Two dual-model structures were developed based on VAE and WGAN.Experimental results indicate that,compared to single-model structures,the dual-model-based VAE and WGAN improve classification accuracy on the test sets by 4.5%and 6%,respectively.

关 键 词：产品质量 异常检测 变分自编码器 Wasserstein生成对抗网络 双模型结构 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]