【作者】 王亚轩；

【导师】 谭峰；

【作者基本信息】 黑龙江八一农垦大学 ， 农业机械化工程， 2022， 博士

【摘要】 我国是大米的生产和消费大国,大米的种植与产地因素有密切关系,特定地域的大米量少质优,不法商贩经常以次充好,消费者在购买时很难判断大米的真实产地。为保护大米的地理特征、维护消费者权益,目前亟需有效的技术手段和方法进行大米的产地真实性鉴别研究。本研究以黑龙江省四个不同产地、相同品种的粳米为研究对象,利用拉曼光谱技术采集大米原始光谱,结合不同预处理方法对原始光谱进行校正和基线去除,利用连续投影算法（SPA）、竞争性自适应重加权算法（CARS）、主成分分析（PCA）的特征提取方法,筛选出全波段和部分波段的大米产地特征波长作为模型输入值,建立了基于偏最小二乘法（PLS）和BP神经网络的大米产地鉴别模型,构建了基于BP神经网络的大米产地鉴别系统。主要研究内容和结论如下:（1）针对产地因素与大米内部成分之间的内在联系,探索了大米中的淀粉、蛋白质和脂肪酸的官能团拉曼光谱解析,并阐述了产地因素对拉曼光谱“产地指纹特征”的作用机理。结果表明,通过对大米的地形地貌、温度、水分、光照、土壤等产地因素分析说明了不同产地的大米在成分上的差异性,利用拉曼光谱对单粒大米中的淀粉、蛋白质和脂肪酸进行了官能团解析,分析得到大米的拉曼谱峰的属性对应为淀粉、蛋白质和油脂,表现形式是相应的分子基团的伸缩振动和弯曲振动。（2）针对大米产地拉曼原始光谱基线漂移问题,提出了一种改进的分段拟合+去除基线的预处理方法,从评价参数和识别正确率比较分析的结果最优。结果表明,拉曼光谱最佳试验参数为积分时间设置为4s、激光强度设置为high、扫描次数为10次、扫描位置为大米中间部位。在此参数下对六种不同产地预处理方法的比较,包括一阶导数+平移平滑、二阶导数+平移平滑、小波变换+去除基线、传统多项式拟合+去除基线、改进的分段拟合+去除基线,最终从评价参数和正确识别率比较分析发现改进的分段拟合+去除基线的结果最优。（3）针对大米产地模型中光谱数据的重叠冗余且不能突出产地特征波长的问题,以黑龙江省绥化、大庆、佳木斯、齐齐哈尔四个不同研究区的龙粳31品种大米为研究对象,利用SPA、CARS、PCA算法提取大米光谱中的特征波长,并将提取的特征波长与大米主要营养物质的特征波长比较分析。利用SPA方法在全波段200-3400cm^-1区间通过变量个数设置分别提取了46个波长点和11个波长点、在部分波段400-1600cm^-1区间分别提取了46个波长点和17个波长点;利用CARS方法在全波段200-3400cm^-1区间提取了100个波长点、在部分波段400-1600cm^-1区间提取了39个波长点;利用PCA方法在400-1600cm^-1&2800-3400cm^-1波段区间通过阈值设置方法提取了8个波长点。结果表明,不同特征方法在400-1600cm^-1&2800-3400cm^-1波段区间提取的波长点,可以落在大米特征峰值或左右肩的上升段及下降段,但在200-400cm^-1&1600-2800cm^-1波段区间提取的波长点为光谱噪声或机器噪声。其中采用PCA方法对大米400-1600cm^-1&2800-3400cm^-1进行特征波长提取,采用设置阈值方法提取大米产地特征谱峰,按照峰值大小顺序提取前四个主成分PC1、PC2、PC3、PC4对应的载荷系数,符合条件的特征谱峰对应20个波长点。最后选取476、867、940、1121、1342、1384、1462和2914cm^-1共计8个特征波长,与大米中主要官能团的特征谱峰解析相符。（4）采用SPA、CARS、PCA方法提取特征波长为大米产地鉴别模型的输入,大米的产地作为输出,建立了基于PLS和BP神经网络法的大米产地鉴别模型,并在目标产地和单一产地、目标产地和混合产地中分别验证了模型的有效性。结果表明:在目标产地和单一产地的模型中PLS模型效果最好为CARS的100点,其正确率是96.9%。BP模型测试集效果最好的是SPA模型部分波段的46个点,其正确率为99.23%;在目标产地和混合产地的模型中,PLS模型效果最好为CARS法100个点,其正确率为95.5%。BP模型测试集效果最好的是部分波段SPA的46个点,其正确率为99.23%。从以上的分析看,BP模型的识别正确率总体高于PLS模型,且在BP模型中SPA部分波段的46个点的输入效果最好,将其选择作为产地鉴别系统的模型输入。（5）针对应用拉曼光谱快速无损检测大米产地问题,构建了基于BP神经网络的在线大米产地鉴别系统。在拉曼光谱现场采集中,通过对拉曼光谱实时采集程序设计、数据库设计、计算方法设计等实现了大米目标产地的快速预测。结果表明:利用该系统对五常产地和五常外其他产地大米拉曼光谱进行测试,采用46个产地特征建模,148个大米样本的鉴别准确率为91.29%。说明大米产地鉴别系统识别准确,操作快捷,可以作为科研人员和广大农业生产单位开展大米产地的鉴别系统。更多还原

【Abstract】 China is a big country of rice production and consumption.Rice cultivation is closely related to the factors of producing area.The quantity of rice in a specific area is small and of high quality.In order to protect the geographical characteristics of rice and protect the rights and interests of consumers,effective technical means and methods are urgently needed to study the authenticity of rice origin.This study four different origin,the same kinds of rice in Heilongjiang province as the research object,the original spectrum of rice was collected by Raman spectroscopy using the original spectrum and combining different pretreatment methods to baseline correction and remove the original spectrum.Using SPA,CARS and PCA feature extraction methods,the characteristic wavelengths of rice producing areas in whole band and part band were selected as the input values of the model.The rice origin identification model based on PLS and BP neural network was established,and a rice origin identification system based on BP neural network was developed.The main research contents and conclusions are as follows:（1）In this paper,the functional group Raman spectroscopy analysis of starch,protein and fatty acid in rice was explored,and the mechanism of origin factors on"origin fingerprint characteristics"of Raman spectrum was expounded.Results show that by topography and geomorphology,the rice production factors such as temperature,moisture,illumination,soil analysis illustrates the different origin of the differences in composition of rice.Raman spectroscopy was used to analyze the functional groups of starch,protein and fatty acid in single grain rice.The properties of Raman spectrum peaks of rice were obtained as starch,protein and fat,and the manifestations were stretching and bending vibrations of corresponding molecular groups.（2）In order to solve the problem of baseline drift of Raman original spectrum in rice producing area,an improved piecewise fitting+baseline removal pretreatment method was proposed,which obtained the best result from the comparison of evaluation parameters and recognition accuracy.The results show that the best experimental parameters of Raman spectrum are as follows:integrating time is set to 4s,laser intensity is set to high,scanning times is set to10,scanning position is set to the middle part of rice.Under the parameters for the comparison of pretreatment methods,six different regions including first derivative+translation smoothing,second derivative+translation smoothing,wavelet transform+baseline removal,traditional polynomial fitting+baseline removal,improved piecewise fitting+baseline removal.Finally from the evaluation parameters and correct recognition rate comparison analysis found that the improved piecewise fitting+baseline removal result is the best.（3）Aiming at the problems of overlapping and redundancy of spectral data in rice producing area model and not being able to highlight the characteristic wavelength of producing area,taking Longjing 31 rice from Suihua,Daqing,Jiamasi and Qiqihar of Heilongjiang Province as the research object,SPA,CARS and PCA algorithms were used to extract the characteristic wavelength of rice spectrum,and the extracted characteristic wavelengths were compared with those of the main nutrients in rice.The SPA method was used to extract 46wavelength points and 11 wavelength points in the whole band 200-3400cm^-1 range and 46wavelength points and 17 wavelength points in the partial band 400-1600cm^-1 range by setting the number of variables.The CARS method was used to extract 100 wavelength points in200-3400cm^-1 and 39 wavelength points in 400-1600cm^-1.PCA method was used to extract 8wavelength points in the band range of 400-1600cm^-1 and 2800-3400cm^-1 by threshold setting method.The results showed that the wavelength points extracted by different feature methods in the band range of 400-1600cm^-1&2800-3400cm^-1 could fall on the peak of rice feature or the ascending and descending segment of left and right shoulder,but the wavelength points extracted in the band range of 200-400cm^-1&1600-2800cm^-1 were spectral noise or machine noise.PCA method was used to extract the characteristic wavelength of rice from 400-1600cm^-1 to2800-3400cm^-1,threshold method was used to extract the characteristic spectrum peak of rice producing area,and the corresponding load coefficients of the four principal components PC1,PC2,PC3 and PC4 were extracted according to the order of peak value.There are 20 wavelength points corresponding to the characteristic spectrum peaks that meet the conditions.Finally,a total of 8 characteristic wavelengths of 476,867,940,1121,1342,1384,1462 and 2914cm^-1were selected,and the main corresponding rice components were starch Raman characteristic spectrum peaks,which were consistent with the analysis of characteristic spectrum peaks of main functional groups in rice.（4）Using SPA,CARS and PCA methods to extract characteristic wavelengths as the input of the rice origin identification model,and the rice origin as the output,the rice origin identification model based on PLS and BP neural network method was established,and the validity of the model was verified in the target origin and single origin,target origin and mixed origin.The results showed that the best PLS model was 100 points of CARS,and the accuracy rate was96.9%.BP model test set has the best effect on 46 points in some bands of SPA model,and its accuracy is 99.23%.In the model of target and mixed origin,PLS model is the best with 100points of CARS method,and its accuracy rate is 95.5%.BP model test set has the best effect on46 points of partial band SPA,and its accuracy rate is 99.23%.From the above analysis,the recognition accuracy of BP model is generally higher than that of PLS model,and the input effect of 46 points in SPA band in BP model is the best,which is selected as the model input of origin identification system database.（5）An on-line rice origin identification system based on BP neural network was constructed to solve the problem of rapid non-destructive detection of rice origin by Raman spectroscopy.In the field acquisition of Raman spectrum,the fast prediction of rice target producing area was realized through the design of real-time acquisition program,database and calculation method of Raman spectrum.The results showed that the rice Raman spectrum of Wuchang and other producing areas was tested by this device,and the identification accuracy of 148 rice spectra was91.29%by using the feature modeling of 46 producing areas.It shows that the rice origin identification system is accurate and quick to operate,and can be used as a scientific research personnel and agricultural production units to carry out rice origin identification.更多还原