Andrade Letícia, Farhat Imad A, Aeberhardt Kasia, Bro Rasmus, Engelsen Søren Balling
University of Nottingham, School of Biosciences, Division of Food Sciences, Sutton Bonington Campus, Loughborough, LE12 5RD, UK.
Appl Spectrosc. 2009 Feb;63(2):141-52. doi: 10.1366/000370209787392094.
The influence of temperature on near-infrared (NIR) and nuclear magnetic resonance (NMR) spectroscopy complicates the industrial applications of both spectroscopic methods. The focus of this study is to analyze and model the effect of temperature variation on NIR spectra and NMR relaxation data. Different multivariate methods were tested for constructing robust prediction models based on NIR and NMR data acquired at various temperatures. Data were acquired on model spray-dried limonene systems at five temperatures in the range from 20 degrees C to 60 degrees C and partial least squares (PLS) regression models were computed for limonene and water predictions. The predictive ability of the models computed on the NIR spectra (acquired at various temperatures) improved significantly when data were preprocessed using extended inverted signal correction (EISC). The average PLS regression prediction error was reduced to 0.2%, corresponding to 1.9% and 3.4% of the full range of limonene and water reference values, respectively. The removal of variation induced by temperature prior to calibration, by direct orthogonalization (DO), slightly enhanced the predictive ability of the models based on NMR data. Bilinear PLS models, with implicit inclusion of the temperature, enabled limonene and water predictions by NMR with an error of 0.3% (corresponding to 2.8% and 7.0% of the full range of limonene and water). For NMR, and in contrast to the NIR results, modeling the data using multi-way N-PLS improved the models' performance. N-PLS models, in which temperature was included as an extra variable, enabled more accurate prediction, especially for limonene (prediction error was reduced to 0.2%). Overall, this study proved that it is possible to develop models for limonene and water content prediction based on NIR and NMR data, independent of the measurement temperature.
温度对近红外（NIR）光谱和核磁共振（NMR）光谱的影响使得这两种光谱方法在工业应用中变得复杂。本研究的重点是分析温度变化对NIR光谱和NMR弛豫数据的影响并进行建模。基于在不同温度下采集的NIR和NMR数据，测试了不同的多元方法来构建稳健的预测模型。在20摄氏度至60摄氏度范围内的五个温度下，对模型喷雾干燥柠檬烯系统的数据进行了采集，并计算了柠檬烯和水预测的偏最小二乘（PLS）回归模型。当使用扩展反演信号校正（EISC）对数据进行预处理时，基于在不同温度下采集的NIR光谱计算的模型的预测能力显著提高。PLS回归平均预测误差降至0.2%，分别相当于柠檬烯和水参考值全范围的1.9%和3.4%。在校准前通过直接正交化（DO）去除温度引起的变化，略微提高了基于NMR数据的模型的预测能力。隐式包含温度的双线性PLS模型能够通过NMR对柠檬烯和水进行预测，误差为0.3%（相当于柠檬烯和水全范围的2.8%和7.0%）。对于NMR，与NIR结果相反，使用多向N-PLS对数据进行建模提高了模型的性能。将温度作为额外变量包含在内的N-PLS模型能够进行更准确的预测，尤其是对于柠檬烯（预测误差降至0.2%）。总体而言，本研究证明可以基于NIR和NMR数据开发柠檬烯和水含量预测模型，而与测量温度无关。
