多判别器生成对抗网络工业不平衡数据建模方法  

Multi-discriminator generative adversarial network industrial imbalance data modeling approach

作　　者：赵佳 杨澜 刘勤学 ZHAO Jia;YANG Lan;LIU Qinxue(College of Science,Hebei University of Technology,Tianjin 300401,China;Ningxia Jiyuan Recycling Development Co.,Ltd.,Shizuishan 753400,China)

机构地区：[1]河北工业大学理学院,天津300401 [2]宁夏吉元循环发展有限公司,宁夏石嘴山753400

出　　处：《计算机集成制造系统》2025年第2期554-566,共13页Computer Integrated Manufacturing Systems

基　　金：国家自然科学基金资助项目(12001153)。

摘　　要：为解决工业场景下不平衡数据建模预测精度较低的问题,提出结合多判别器生成对抗网络及反聚类筛选的工业不平衡数据建模方法来增强模型分类预测效果。针对生成对抗网络模型在训练过程中存在模式崩溃,导致生成数据多样性差的问题,基于集成思想,使用多判别器框架改进Wasserstein生成对抗网络,增强模型对模式崩溃问题的鲁棒性;针对生成数据存在噪声的问题,集成有序点集识别聚类结构算法和高斯混合模型聚类算法从密度及分布角度对生成数据进行聚类,采用信息熵反向筛选生成数据扩充少数类样本;在电极升降数据集及UCL带钢缺陷数据集上采用XGBOOST、支持向量机、BP神经网络3种分类模型对比原始不平衡数据、随机过采样、SMOTE算法、原始生成对抗网络与所提方法解决不平衡问题后模型的分类预测效果。实验验证了所提方法的优越性。To solve the problem of low prediction accuracy of imbalance data modeling in industrial scenarios,an industrial imbalance data modeling method combining multi-discriminator generative adversarial network and inverse clustering screening was proposed to enhance the model classification prediction effect.Aiming at the problem of pattern collapse in training process of generative adversarial network models that leaded to poor diversity of generated data,based on the integration idea,Wasserstein generative adversarial network was improved using a multi-discriminator framework to enhance the robustness of the model to the pattern collapse problem.For the problem of noise in the generated data,integrating OPTICS and GMM clustering model to cluster the generated data from the perspective of density and distribution,and using information entropy to reverse screen the generated data to expand a small number of class samples.XGBOOST,SVM and BP neural network three classification models were used to compare the classification prediction effect of the models after solving the imbalance problem with the original imbalance data,random oversampling,SMOTE algorithm,GAN and the proposed method on the electrode lifting and lowering dataset and the UCL steel plates faults dataset.Experiments showed that the method could effectively improve the Recall,Precision and F1 values,enhance the sensitivity and comprehensiveness of the identification of the state of the mineral heat furnace and strip defects,ensure the safe operation of the mineral heat furnace,reduce the number of low-quality steel and defective strips,and improve product quality.

关 键 词：工业不平衡数据 生成对抗网络 生成数据筛选 信息熵 矿热炉 电极升降 带钢缺陷识别 

分 类 号：TP311.13[自动化与计算机技术—计算机软件与理论]