一维卷积神经网络的手持式可见/近红外柑橘可溶性固形物含量无损检测系统  

Hand-Held Visible/Near Infrared Nondestructive Detection System for Soluble Solid Content in Mandarin by 1D-CNN Model

作　　者：蔡健荣[1,2] 黄楚钧 马立鑫 翟利祥 郭志明 CAI Jian-rong;HUANG Chu-jun;MA Li-xin;ZHAI Li-xiang;GUO Zhi-ming(School of Food and Biological Engineering,Jiangsu University,Zhenjiang 212013,China;International Joint Research Laboratory of Intelligent Agriculture and Agri-Products Processing(Jiangsu University),Jiangsu Education Department,Zhenjiang 212013,China;Key Laboratory of Modern Agricultural Equipment and Technology(Jiangsu University),Ministry of Education,Zhenjiang 212013,China)

机构地区：[1]江苏大学食品与生物工程学院,江苏镇江212013 [2]江苏省智能农业与农产品加工国际合作联合实验室,江苏镇江212013 [3]现代农业装备与技术教育部重点实验室(江苏大学),江苏镇江212013

出　　处：《光谱学与光谱分析》2023年第9期2792-2798,共7页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(51975259),财政部和农业农村部:国家现代农业产业技术体系项目,江苏高校青蓝工程项目,现代农业装备与技术教育部重点实验室开放基金项目(MAET202117),江苏大学农业装备学部青年项目(NZXB20210205)资助。

摘　　要：为实现柑橘可溶性固形物含量(SSC)快速无损检测,基于可见/近红外技术开发了低功耗手持式柑橘可溶性固形物含量无损检测系统。以宽谱LED光源结合特征窄带微型光谱仪为核心,设计了手持式柑橘可溶性固形物含量无损检测终端。开发了基于物联网技术的水果光谱仪云端数据系统,该系统主要包括用户库、设备库、检测数据库和模型库,通过通讯模块与手持式无损检测终端相连接,可以实现光谱采集参数修改、云端数据上传与下载、云模型的调用等功能。利用该检测系统获取的光谱数据,建立一维卷积神经网络(1D-CNN)模型用于预测柑橘的可溶性固形物含量。该网络包含输入层、卷积层、池化层、全连接层和输出层等7层结构。主机采集柑橘的光谱数据并建立1D-CNN柑橘可溶性固形物含量预测模型,并用该模型与多种传统回归方法进行对比。1D-CNN模型的预测相关系数和预测均方根误差分别为0.812,0.488,优于偏最小二乘法(PLS),人工神经网络(ANN)和支持向量机(SVM)。采用基于模型的迁移学习方法,基于主机的1D-CNN模型对从机进行模型传递,研究了从机标准样本数量对模型传递的影响。发现使用少量从机光谱样本即可取得较好的效果,从机预测集均方根误差为0.531。研究结果表明,研发的柑橘SSC云模型的手持式可见近红外无损检测系统具有检测快速、低成本、操作简便等优点,基于该检测系统的1D-CNN网络可以有效提取柑橘光谱的有效特征并进行回归分析。借助迁移学习算法,可以实现1D-CNN模型在不同装置间的有效传递,满足柑橘可溶性固形物含量无损检测的需求。为手持式水果内部品质无损检测系统的开发与应用提供了借鉴和参考。To realize the rapid,nondestructive detection of solid soluble content(SSC)in Mandarin,a hand-held nondestructive detection system was developed based on visible/near-infrared technology.Wide spectra range LED light source combined with a narrowband response micro-spectrometer was designed as a core in the handheld nondestructive detection terminal.The cloud data system of the fruit spectrometer based on Internet of Things technology was also developed,including a user database,equipment database,test database and model database.The data system was connected with the detection terminal through a communication module to realize functions,including modifying parameters of spectra collection,uploading and downloading the cloud data and invoking cloud model.Based on the spectra collected by the system,anovelone-dimensional convolutional neural network(1D-CNN)model was proposed to predictmandarin soluble solid content.The network contains 7 layers:input,convolution,pooling,full connection,and output.Mandarin spectra of the master machine were collected to build the 1D-CNN SSC prediction model,and the 1D-CNN model was compared with traditional regression methods to evaluate the model performance.The R p and RMSEP of the 1D-CNN model were 0.812 and 0.488 respectively,better than that of partial least squares(PLS),artificial neural network(ANN)and support vector machine(SVM).Transfer learning method based on the 1D-CNN model of the master machine was adopted to transfer the model to the slave machine,the influence of the number of samples from the slave machine on model transfer was studied,and a small number of slave machine spectral samples for model training achieved good model transfer effect,modeling transferring result with root mean square error of prediction of the slave machine being 0.531.The results demonstrated that the detection system has the advantages of fast detection,low cost and simple operation.The 1D-CNN network based on the detection system could effectively extract effective features of Mandarin sp

关 键 词：无损检测 柑橘 可见/近红外光谱 可溶性固形物含量 一维卷积神经网络 迁移学习 模型传递 

分 类 号：O657.3[理学—分析化学]