一种生成对抗网络半监督回归的废纺样品中羊毛含量的预测方法  

Prediction Method of Wool Content in Waste Spinning Samples Based on Semi Supervised Regression of Generative Adversarial Network

作　　者：胡锦泉 杨辉华[1] 赵国樑 周瑞知 李灵巧 HU Jin-quan;YANG Hui-hua;ZHAO Guo-liang;ZHOU Rui-zhi;LI Ling-qiao(School of Artificial Intelligence,Beijing University of Posts and Telecommunications,Beijing 100876,China;School of Electronic Engineering and Automation Chemistry,Guilin University of Electronic Technology,Guilin 541004,China;Beijing Institure of Fashion Technology,Beijing 100029,China;Shanghai Jicai Environmental Protection Technology Co.,Ltd.,Shanghai 200131,China;School of Computer Science and Information Security,Guilin University of Electronic Technology,Guilin 541004,China)

机构地区：[1]北京邮电大学人工智能学院,北京100876 [2]桂林电子科技大学电子工程与自动化学院,广西桂林541004 [3]北京服装科技学院,北京100029 [4]上海季采环保科技有限公司,上海200131 [5]桂林电子科技大学计算机学院,广西桂林541004

出　　处：《光谱学与光谱分析》2024年第5期1417-1424,共8页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(62262010);广西自动检测技术与仪器重点实验室项目(PF18078X)资助。

摘　　要：针对废旧纺织品回收在线分拣的需求,提出了一种基于生成对抗网络的半监督回归方法,使用少量标记样本和大量未标记样本来训练半监督回归器。半监督回归器由神经网络组成的生成器和由神经网络构成的判别器组成。生成器用于生成尽可能接近实际标记和未标记训练数据集内容的混合样本。鉴别器用于验证生成器生成的样本并预测这些样本的连续标记。生成的网络通过特征匹配损失进行训练,损失函数是鉴别器中间层真实样本的输出与生成样本之间的误差平均值。判别式有两个输出,一个用于预测序列标记,另一个用于确定生成的样本是真样本还是假样本的概率。通过使用传统的无监督生成对抗性网络损失函数和监督回归损失的组合来训练判别式。生成的网络通过特征匹配损失进行训练,损失函数是鉴别器中间层真实样本的输出与生成样本之间的误差平均值。先后收集了400个不同羊毛含量的混纺样品和3000个未知成分的混纺样品。70%的标记和未标记的混合样本被随机选择作为训练集,其余30%的标记样本被用作重复实验的测试集。开展了多个实验进行验证。第一个实验为混纺光谱生成实验,用于验证生成对抗网络能够根据内在规律有效生成混合样本光谱。第二个实验为半监督对抗网络定量分析性能对比实验,对羊毛成分分析模型训练与测试,并将本半监督对抗网络定量分析模型与其他定量模型进行性能对比。第三个实验为现场高羊毛含量混纺细分模型预测对比实验,用羊毛含量在80%到99%之间的混纺样品进行成分分析,并将本文的半监督对抗网络定量分析模型与其他定量模型进行性能对比。第四个实验为中高羊毛含量混纺细分综合模型现场预测实验,用羊毛含量在40%到99%之间的混纺样品训练半监督对抗网络定量分析模型并部署在分拣系统,由操�In this paper,a semi-supervised regression method based on a Generative adversarial network is proposed to meet the demand for online sorting for waste textile recycling,which uses some labeled samples and a large number of unlabeled samples to train the semi-supervised regression.The semi-supervised regression consists of a generator composed of neural networks and a discriminator composed of neural networks.The generator generates a mixed sample that is as close as possible to the actual labeled and unlabeled training dataset content.The discriminator is used to validate the samples generated by the generator and predict the continuous labeling of these samples.The generated network is trained through feature matching loss,which is the average error between the output of the actual sample in the middle layer of the discriminator and the generated sample.The discriminant has two outputs,one for predicting sequence markers and the other for determining the probability of whether the generated sample is a true or false sample.Train discriminants by using a combination of traditional unsupervised generative adversarial network loss functions and supervised regression losses.The generated network is trained through feature matching loss,which is the average error between the output of the actual sample in the middle layer of the discriminator and the generated sample.We have collected 400 blended samples with different wool contents and 3000 blended samples with unknown components.70%of labeled and unlabeled mixed samples were randomly selected as the training set,while the remaining 30%of labeled samples were used as the test set for repeated experiments.This article has conducted multiple experiments for verification.The first experiment is a blended spectrum generation experiment,which is used to verify that the generative adversarial network can effectively generate mixed sample spectra based on intrinsic laws.The second experiment is a semi-supervised adversarial network quantitative analysis performance compar

关 键 词：生成对抗网络 废纺回收 近红外光谱 羊毛含量 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]