基于Dropout深度信念网络的棉和涤纶含量的近红外光谱检测  

Near-Infrared Spectroscopy Detection of Cotton/Polyester Content Based on Dropout Deep Belief Network

作　　者：韦卓 李稳稳 林敏 江文松 周新奇 Wei Zhuo;Li Wenwen;Lin Min;Jiang Wensong;Zhou Xinqi(College of Metrology and Measurement Engineeriyig,China Jiliang University,Hangzhou,Zhejiang 310018,China;Hangzhou Puyu Technology Inc.,Hangzhou,Zhejiang 310023,China)

机构地区：[1]中国计量大学计量测试工程学院,浙江杭州310018 [2]杭州谱育科技发展有限公司,浙江杭州310023

出　　处：《光学学报》2021年第17期189-195,共7页Acta Optica Sinica

基　　金：国家重点研发计划项目(2018YFF01012006);国家重大科学仪器设备开发专项(2014YQ470377);浙江省基本科研业务费资助项目(2021YW82)。

摘　　要：将近红外光谱技术与深度学习理论相结合,提出了一种基于Dropout深度信念网络(DBN)的棉涤混纺面料中各组分含量的快速检测方法。首先使用小波变换对原始光谱数据进行压缩处理,再构建以高斯受限玻尔兹曼机(GRBM)为核心的DBN模型,以保证输入数据信息的完整性;然后利用Dropout来防止模型过拟合,通过隐藏部分隐含层节点来减小节点之间的相互依赖,实现网络的稀疏化处理,提高了非线性建模和网络模型的泛化能力。实验结果表明:对于采用Dropout-DBN方法建立的棉涤混纺面料中各组分含量的分析模型,其棉、涤纶含量的预测集相关系数分别为0.9927和0.9903,预测集均方根误差分别为0.0792和0.0869。与其他建模方法相比,所建模型的精度和适应性显著提高,并有利于模型的传递与共享,提高了模型的智能化。We combined near-infrared spectroscopy with the deep learning theory to propose a method for rapidly detecting the content of each component in cotton-polyester blended fabric based on the Dropout deep belief network(DBN).Firstly,wavelet transform was used to compress the original spectral data.Then,a DBN model with a Gaussian restricted Boltzmann machine(GRBM)as the core was constructed,which could ensure the integrity of input data.Finally,Dropout was used to effectively prevent the model from overfitting and the interdependence between nodes was reduced by hiding some hidden layer nodes.As a result,network sparsification was achieved and the ability of nonlinear modeling and network model generalization was enhanced.The experimental results indicate that in the analytical model of the content of each component in cotton-polyester blended fabric built by the Dropout-DBN method,the correlation coefficients of prediction set for cotton and polyester contents are respectively 0.9927 and 0.9903,and the root mean square errors of prediction set are 0.0792 and 0.0869,respectively.The model proposed in this paper has much higher accuracy and adaptability than other modeling methods,which is conducive to model transfer and sharing and improves the intelligentization of the model.

关 键 词：光谱学 近红外光谱 棉涤混纺面料 深度信念网络 高斯受限玻尔兹曼机 Dropout策略 

分 类 号：O657.33[理学—分析化学] TS116[理学—化学]