变量优选补正算法的鲜枣可溶性固形物检测模型传递方法研究
|
摘要
在水果的品质检测和分级分选中,存在不同仪器所建检测模型难以共享的难题。为此,以壶瓶枣为研究对象,利用可见/近红外光谱技术探讨仪器间可溶性固形物含量(SSC)检测模型的传递方法。首先,采用美国ASD(Analytical Spectral Device)公司生产的两台仪器采集样本的光谱信息,采用最小二乘支持向量机(LS-SVM)建立原始光谱、 Savitzky-Golay一阶导数处理、标准正态变量变换后的SSC检测模型,预测不同仪器采集的光谱时3种方法的预测能力均较差。预测同一台仪器的光谱时,基于原始光谱的主仪器所建模型最优,预测集的决定系数(R■)和均方根误差(RMSEP)分别为0.73和1.36%。在此基础上,采用Kennard/Stone算法选取标样,利用专利算法(Shenk’s)、直接标准化(DS)、斜率/偏差算法(S/B)进行模型传递。然后,根据回归系数提取主仪器(24个)和从仪器(28个)的特征波长,优选出单一变量(SV)24个、共性变量(CV)23个、融合变量(FV)29个,均涵盖了SSC的主要吸收谱带。利用优选的变量分别建立主仪器的LS-SVM检测模型,采用主仪器的预测结果(R■=0.78～0.80, RMSEP=1.07%～1.13%)明显好于全波段所建模型,但预测从仪器时RMSEP为6.62%～7.88%,模型失效。最后,基于波长位置偏移和分子振动的吸收特性提出了共性变量优选结合差值补正(CV-MC)、单一变量优选结合差值补正、融合变量优选结合差值补正、共性变量优选结合波长补正算法(CV-WC)进行模型传递,并与SV-Shenk’s, CV-Shenk’s, FV-Shenk’s, SV-DS, CV-DS, FV-DS, SV-S/B, CV-S/B和FV-S/B进行对比分析。结果表明,基于全波段进行模型传递时,预测结果均较差(R■=0.03～0.34, RMSEP=2.44%～4.67%);基于优选变量所建模型经SV-Shenk’s, CV-Shenk’s, FV-Shenk’s传递后的结果较差,经其他算法传递后的结果(R■=0.47～0.73, RMSEP=1.30%～1.90%)好于全波段;基于共性变量传递后的结果好于单一变量和融合变量, CV-MC结果最佳(R~2_p=0.73, RMSEP=1.30%), CV-WC传递后的预测结果(RMSEP=1.62%)与CV-DS和CV-S/B相近。研究表明, CV-MC和CV-WC均是一种有效模型传递算法,对建立不同仪器间通用的鲜枣品质检测模型具有重要意义。
The difficulty of sharing detection models built by different instruments is common in the quality inspection and classification of fruits. In the study, the "Huping" jujube was used as the research object, and the transfer method of the soluble solids content(SSC) detection model between instruments was explored using visible/near infrared spectroscopy. First, the spectral information of samples was collected using two instruments produced by American Analytical Spectral Device. Based on the original, Savitzky-Golay first derivative processed and standard normal variable transformed spectrum, SSC detection models were established by least squares-support vector machines(LS-SVM), respectively. The prediction ability of the three methods for the spectra acquired by different instruments was poor. The built model by the original spectrum of the master instrument was optimal in predicting spectra from the same instrument. The determination coefficient(R■) and the root mean squared error of prediction(RMSEP) were 0.73 and 1.36%, respectively. Next, the Kennard/Stone algorithm was used to select standard samples. The Shenk's, direct standardization(DS) and slope/bias(S/B) algorithm were used for model transfer, respectively. Then, according to the regression coefficient, the sensitive wavelengths of the master instrument(24) and the slave instrument(28) were extracted. 24 single variables(SV), 23 common variables(CV) and 29 fusion variables(FV) were selected, all of which contained the main absorption bands of SSC. LS-SVM detection models of the master instrument were respectively established by the preferred variables, which(R■=0.78～0.80, RMSEP=1.07%～1.13%) was better than the model built by the full wavelength for the prediction result of the master instrument. However, the model failed in predicting spectra from different instruments(RMSEP=6.62%～7.88%). Finally, based on the wavelength position shift and the absorbed property of molecular vibration, these algorithms named as common variable-subtraction correction(CV-MC), single variable-subtraction correction, fusion variable-subtraction correction and common variable-wavelength correction(CV-WC) were respectively proposed for model transfer. These methods were compared with SV-Shenk's, CV-Shenk's, FV-Shenk's, SV-DS, CV-DS, FV-DS, SV-S/B, CV-S/B and FV-S/B algorithms. The results showed that the prediction results(R■=0.03～0.34, RMSEP=2.44%～4.67%) were poor when the model was transferred by the full-band. Using the model built by the preferred variables, the results transferred by SV-Shenk's, CV-Shenk's and FV-Shenk's were poor, and the results transferred by other algorithms(R■=0.47～0.73, RMSEP=1.30%～1.90%) were better than the full wavelength. The CV got better transfer results than the SV and the FV, and the CV-MC result was the best(R■=0.73, RMSEP=1.30%). The predicted result after CV-WC transfer(RMSEP=1.62%) was similar to CV-DS and CV-S/B. The research indicates that both CV-MC and CV-WC are effective model transfer algorithms, which are of great significance to establishing a common jujube quality detection model between different instruments.
The difficulty of sharing detection models built by different instruments is common in the quality inspection and classification of fruits. In the study, the "Huping" jujube was used as the research object, and the transfer method of the soluble solids content(SSC) detection model between instruments was explored using visible/near infrared spectroscopy. First, the spectral information of samples was collected using two instruments produced by American Analytical Spectral Device. Based on the original, Savitzky-Golay first derivative processed and standard normal variable transformed spectrum, SSC detection models were established by least squares-support vector machines(LS-SVM), respectively. The prediction ability of the three methods for the spectra acquired by different instruments was poor. The built model by the original spectrum of the master instrument was optimal in predicting spectra from the same instrument. The determination coefficient(R■) and the root mean squared error of prediction(RMSEP) were 0.73 and 1.36%, respectively. Next, the Kennard/Stone algorithm was used to select standard samples. The Shenk's, direct standardization(DS) and slope/bias(S/B) algorithm were used for model transfer, respectively. Then, according to the regression coefficient, the sensitive wavelengths of the master instrument(24) and the slave instrument(28) were extracted. 24 single variables(SV), 23 common variables(CV) and 29 fusion variables(FV) were selected, all of which contained the main absorption bands of SSC. LS-SVM detection models of the master instrument were respectively established by the preferred variables, which(R■=0.78～0.80, RMSEP=1.07%～1.13%) was better than the model built by the full wavelength for the prediction result of the master instrument. However, the model failed in predicting spectra from different instruments(RMSEP=6.62%～7.88%). Finally, based on the wavelength position shift and the absorbed property of molecular vibration, these algorithms named as common variable-subtraction correction(CV-MC), single variable-subtraction correction, fusion variable-subtraction correction and common variable-wavelength correction(CV-WC) were respectively proposed for model transfer. These methods were compared with SV-Shenk's, CV-Shenk's, FV-Shenk's, SV-DS, CV-DS, FV-DS, SV-S/B, CV-S/B and FV-S/B algorithms. The results showed that the prediction results(R■=0.03～0.34, RMSEP=2.44%～4.67%) were poor when the model was transferred by the full-band. Using the model built by the preferred variables, the results transferred by SV-Shenk's, CV-Shenk's and FV-Shenk's were poor, and the results transferred by other algorithms(R■=0.47～0.73, RMSEP=1.30%～1.90%) were better than the full wavelength. The CV got better transfer results than the SV and the FV, and the CV-MC result was the best(R■=0.73, RMSEP=1.30%). The predicted result after CV-WC transfer(RMSEP=1.62%) was similar to CV-DS and CV-S/B. The research indicates that both CV-MC and CV-WC are effective model transfer algorithms, which are of great significance to establishing a common jujube quality detection model between different instruments.
引文
[1] LI Xin, BIN Jun, FAN Wei, et al(李鑫, 宾俊, 范伟, 等). Chinese Journal of Analytical Chemistry(分析化学), 2017, 45(7): 958.
[2] Qin Yuhua, Gong Huili. Infrared Physics & Technology, 2016, 77: 239.
[3] Mou Yi, Zhou Long, Yu Shujian, et al. Chemometrics and Intelligent Laboratory Systems, 2016, 156: 62.
[4] LIU Jiao, LI Xiao-yu, GUO Xiao-xu, et al(刘娇, 李小昱, 郭小许, 等). Transactions of the Chinese Society of Agricultural Engineering(农业工程学报), 2014, 30(17): 276.
[5] WENG Hai-yong, CEN Hai-yan, HE Yong(翁海勇, 岑海燕, 何勇). Spectroscopy and Spectral Analysis(光谱学与光谱分析), 2018, 38(1): 235.
[6] Liu Yan, Cai Wensheng, Shao Xueguang. Spectrochimica Acta Part A Molecular and Biomolecular Spectroscopy, 2016, 169: 197.
[7] Liang Chen, Yuan Hongfu, Zhao Zhong, et al. Chemometrics and Intelligent Laboratory Systems, 2016, 153: 51.
[8] XU Hui-rong, LI Qing-qing(徐惠荣, 李青青). Transactions of the Chinese Society for Agricultural Machinery(农业机械学报), 2017, 48(9): 312.
[9] Zhang Shujuan, Zhang Haihong, Zhao Yanru, et al. Mathematical and Computer Modelling, 2013, 58(3-4): 545.
[10] Kennard R W, Stone L A. Technometrics, 1969, 11(1): 137.
[11] Yao Yang, Chen Huarui, Xie Lijuan, et al. Journal of Food Engineering, 2013, 119(1): 22.
[12] NI Li-jun, HAN Ming-yue, ZHANG Li-guo, et al(倪力军, 韩明月, 张立国). Chinese Journal of Analytical Chemistry(分析化学), 2018, 46(10): 1660.
[13] Travers S, Bertelsen M G, Petersen K K, et al. LWT-Food Science and Technology, 2014, 59(2): 1107.
[14] Jerry Workman Jr, Lois Weyer. Practical Guide to Interpretive Near-Infrared Spectroscopy(近红外光谱解析实用指南). Translated by CHU Xiao-li, XU Yu-peng, TIAN Gao-you(褚小立, 许育鹏, 田高友, 译). Beijing: Chemical Industry Press(北京: 化学工业出版社), 2009. 19.
[2] Qin Yuhua, Gong Huili. Infrared Physics & Technology, 2016, 77: 239.
[3] Mou Yi, Zhou Long, Yu Shujian, et al. Chemometrics and Intelligent Laboratory Systems, 2016, 156: 62.
[4] LIU Jiao, LI Xiao-yu, GUO Xiao-xu, et al(刘娇, 李小昱, 郭小许, 等). Transactions of the Chinese Society of Agricultural Engineering(农业工程学报), 2014, 30(17): 276.
[5] WENG Hai-yong, CEN Hai-yan, HE Yong(翁海勇, 岑海燕, 何勇). Spectroscopy and Spectral Analysis(光谱学与光谱分析), 2018, 38(1): 235.
[6] Liu Yan, Cai Wensheng, Shao Xueguang. Spectrochimica Acta Part A Molecular and Biomolecular Spectroscopy, 2016, 169: 197.
[7] Liang Chen, Yuan Hongfu, Zhao Zhong, et al. Chemometrics and Intelligent Laboratory Systems, 2016, 153: 51.
[8] XU Hui-rong, LI Qing-qing(徐惠荣, 李青青). Transactions of the Chinese Society for Agricultural Machinery(农业机械学报), 2017, 48(9): 312.
[9] Zhang Shujuan, Zhang Haihong, Zhao Yanru, et al. Mathematical and Computer Modelling, 2013, 58(3-4): 545.
[10] Kennard R W, Stone L A. Technometrics, 1969, 11(1): 137.
[11] Yao Yang, Chen Huarui, Xie Lijuan, et al. Journal of Food Engineering, 2013, 119(1): 22.
[12] NI Li-jun, HAN Ming-yue, ZHANG Li-guo, et al(倪力军, 韩明月, 张立国). Chinese Journal of Analytical Chemistry(分析化学), 2018, 46(10): 1660.
[13] Travers S, Bertelsen M G, Petersen K K, et al. LWT-Food Science and Technology, 2014, 59(2): 1107.
[14] Jerry Workman Jr, Lois Weyer. Practical Guide to Interpretive Near-Infrared Spectroscopy(近红外光谱解析实用指南). Translated by CHU Xiao-li, XU Yu-peng, TIAN Gao-you(褚小立, 许育鹏, 田高友, 译). Beijing: Chemical Industry Press(北京: 化学工业出版社), 2009. 19.