由上海棱光技术有限公司与中国农业大学联合研发的S450型近红外高密度光栅光谱仪, 使用高速采集技术可得到高密度光谱(波长范围9002 500 nm, 采集间隔0.1 nm, 光谱包含16 001个数据点), 本文采用该仪器并以小麦、 烟草样品为实验对象, 针对高密度光谱的数据特点, 采用S.G.(savitzky-golay)平滑、 固定窗口组合滑动窗口平滑(FCMWS和一阶导数(FD)等数据处理方法, 并应用偏最小二乘法(PLS)对小麦粗蛋白、 烟草烟碱及总糖含量进行建模和预测, 对仪器整体性能以及数据处理方法的参数优化等, 进行了评价和比较研究。 结果表明: (1)小麦、 烟草样品的原光谱经S.G.平滑结合一阶导数预处理后, 模型性能大幅提高, 通过对参数拟合阶次 M和平滑点数 N进行优化得出, 当 M一定时, N可选取范围较宽, 且当 M=2和N处于201~801区间时模型效果理想且稳定; (2)FCMWS方法对小麦、 烟草样品的原光谱进行两层平均平滑, 经调整优化平滑参数 K₁和 K₂( K₁为第一层平滑的固定窗口大小, K₂为第二层滑动窗口大小)得出: 两层平滑参数相乘约为150310时, 模型性能稳定且较优, 同时FCMWS方法极大地提高了建模速度; (3)以小麦样品为对象, 同时在两台S450型光谱仪上采集样品光谱, 对比分析了仪器间的性能差异, 结果表明光谱经S.G.平滑或FCMWS方法处理后, 不同仪器模型间相互预测数据的相对偏差小于2.00%, 远低于预测值与参考值间的相对偏差, 说明上述两种方法均可降低仪器的台间差异, 实现台间模型的稳定传递。 研究结果表明, 国产S450型高密度光栅光谱仪结合数据平滑去噪技术, 已满足小麦、 烟草等农产品品质检测和模型传递的性能要求, 且该光栅型仪器成本相对较低, 对农业领域推广近红外快速检测技术的应用具有实际意义。

In this paper, we used the S450 near-infrared high-density grating spectrometer with technology of high-speed acquisition developed by Shanghai Lengguang Technology Co., Ltd. and China Agricultural University, took wheat and tobacco as the experimental object, and aimed at the high-density spectra (wavelength range is 900 2500nm, interval of wavelength is 0.1 nm, contains 16 001 data points). By adapting processing methods such as S.G. (Savitzky-Golay) smooth, FCMWS (Fixed window combine moved window smoothing) and the First Derivative, Partial Least Squares (PLS) was also used to model and predict the content of crude protein in wheat, nicotine and total sugar in tobacco, evaluate performance of the spectrometer, and optimize the parameters of processing methods. The results show that: (1) The performance of the models was greatly improved after the high density spectrum was processed by S.G. and the first derivative. Optimizing the parameter M (fitting order) and N(number of smoothing point) , if M is a fixed number, N can be selected from a wider range, and when M=2, N is in the interval of 201801, the performance of models is ideal and stable; (2) The FCMWS was designed for smoothing layers of two, fixed window size of the first layer K₁ and second layer K₂ , and it was concluded that the performance of models is better and superior when the multiplication of K₁ and K₂ is about 150310, moreover the FCMWS algorithm is speedy in modeling. (3) In order to analyze instrument differences, only took wheat as the object, which was measured by two S450 spectrometers, experimentally, whether the spectrum is processed by S.G. or FCMWS, the relative deviation of the predicted data from different models between instruments is less than 2.00%, which is far lower than the relative deviation between the predicted and reference values. It indicates that the above two methods can reduce the instrument differences and models can transfer stably among instruments. For wheat, tobacco and other agricultural products, the results of this study reflect that the domestic high-density grating spectrometer S450 combined with de-noising methods, can meet the actual requirements of quality detection and model transfer, and the grating instrument is relatively low-cost, which is significant for popularizing application of the rapid detection technology of near infrared in the agricultural field.