针对不同型号的近红外光谱仪器(主机: SupNIR-2700, 从机: Nicolet Antaris Ⅱ)间的模型传递和同一仪器(Nicolet Antaris Ⅱ)不同分辨率的光谱间的模型传递进行研究, 提出了一种改进的PDS算法-SP-SG1st-PDS算法, 该方法结合三次样条插值、 Savitaky-Golay一阶求导和PDS算法。 思路是通过三次样条插值拟在不破坏原始光谱固有的信息的前提下实现了主光谱与从光谱之间的匹配, 然后对光谱进行S-G一阶求导去除光谱的基线漂移, 再通过PDS算法进行模型传递, 有效消除主、 从光谱之间的差异, 提高多元校正模型的预测精度。 该方法用于醋酸乙烯酯在乙烯-乙酸乙烯酯共聚物中含量的研究, 并且与小波去噪方法和S-G平滑方法作比较。 实验表明: 对于不同型号的仪器间的模型传递, 新方法采用S-G一阶求导较其他方法有明显的优势, 其验证集预测精密度RMSEP从20.595 0降低至0.374 8, 明显优于S-G平滑(0.522 1)和小波去噪(0.516 7)方法, 预测偏差也同样地被改善。 对于同一仪器不同分辨率的光谱之间的模型传递, 在模型传递前后其模型预测精密度RMSEP从0.272 2减少至0.255 3。 通过提出的SP-SG1st-PDS算法, 模型传递能应用于不同类型仪器之间, 也能用于相同仪器不同分辨率的光谱之间, 并且取得了满意的传递结果。

Although, near-infrared (NIR) has been successfully applied to rapid and non-destructive analysis in various fields, however the problem that calibration model developed on one instrument can’t be directly applied to other instruments in many cases has not been solved yet. To achieve calibration model transfer between different types of instruments (master instrument: SupNIR-2700, slave instrument: Nicolet AntarisⅡ) and the model transfer between the spectra of different resolutions measured on one instrument (Nicolet AntarisⅡ), an improved calibration model transfer method is proposed in this work, being referred as SP-SG1st-PDS method. In terms of the proposed method, firstly, a fitting slave spectrum is constructed with cubic spline interpolation without destroying the information on the original slave spectrum. Then, Savitaky-Golay (first order derivative) smoothingis applied to remove the baseline drift between the master and slave spectra. Finally, PDS is applied to the model transfer to remove most of the differences between the master and slave spectra. SP-SG1st-PDS method has been applied to predict the content of vinyl acetate (VAC) in ethylene-vinyl acetate copolymer (EVA). Comparative studies of the proposed model transfer technique, the wavelet de-noising method (SP-WT-PDS) and the S-G smoothing method (SP-SG-PDS) have also been accomplished. As for the model transfer between different types of instrument, comparative experiment results have shown thatthe root-mean-square error of prediction (RMSEP) with the proposed SP-SG1st-PDS method reduced from 15.978 2 to 0.239 0 and much smaller than that with SP-SG-PDSmethod (0.549 0) and that with SP-WT-PDSmethod (0.528 8). Meanwhile, the bias of prediction was improved obviously after the model transfer with the proposed method. For the model transfer between the spectra at different resolutions measured on one instrument, the comparative experiment results have shown that the model prediction accuracy can be improved significantly, accompanied by the reduction in RMSEP from 0.475 1 to 0.194 5. With the proposed SP-SG1st-PDS method, the calibration model transfer can be applied to different types of instruments and different resolutions of spectra measured on the same instrument.