College of Engineering and Technology, Tianjin Agricultural University, Tianjin 300384, China.
College of Engineering and Technology, Tianjin Agricultural University, Tianjin 300384, China.
Spectrochim Acta A Mol Biomol Spectrosc. 2022 Apr 15;271:120958. doi: 10.1016/j.saa.2022.120958. Epub 2022 Jan 29.
To improve the robustness of near infrared (NIR) identification models for the milk adulteration, a novel approach was explored based on asynchronous two-dimensional correlation spectroscopy (2D-COS) slice spectra obtained at characteristic wavebands for pure milk and adulterant combined with an N-way partial least squares discriminant analysis (NPLS-DA). NIR diffuse reflectance spectra from four different brands, Guangming (GM), Mengniu (MN), Sanyuan (SY), and Wandashan (WDS), were collected in range of 11,000 to 4000 cm. Influence of brands on discrimination models for adulterated milk was analyzed. The asynchronous 2D-COS slice spectra at 10 characteristics wavebands, including 4 wavebands for pure milk and 6 wavebands for urea, were input into NPLS-DA to construct discriminant models. External validations using five independent calibration sets from intrabrand or interbrand were established. The same prediction set of 26 SY samples was used to assess the prediction ability of different calibration sets and compared with traditional one-dimensional (1D) NIR spectra based on a partial least squares discriminant analysis (PLS-DA). It showed that for intrabrand model, the correct rates for the calibration and predication sets were 100% and 96.15%, respectively. For the interbrand model, the correct rates by the NPLS-DA for the calibration set of GM, MN, and WDS milk were both 100%. The corresponding rates for the prediction set were 73%, 88.46% and 69.23%, respectively, which were much higher than those of PLS-DA (only 50%, 53.83% and 50%, respectively). It was proven that model robustness was sensitive to changes in the milk brands. The proposed method can effectively reduce the influence of brands on the discrimination models.
为了提高近红外（NIR）识别模型对牛奶掺假的鲁棒性，探索了一种基于纯牛奶和掺杂物特征波段异步二维相关光谱（2D-COS）切片光谱的新方法，并结合 N 路偏最小二乘判别分析（NPLS-DA）。在 11000 到 4000 cm 的范围内，收集了来自四个不同品牌（光明、蒙牛、三元和完达山）的 NIR 漫反射光谱。分析了品牌对掺假牛奶判别模型的影响。将 10 个特征波段的异步 2D-COS 切片光谱（包括 4 个纯牛奶波段和 6 个尿素波段）输入 NPLS-DA 构建判别模型。建立了来自同一品牌或不同品牌的五个独立校准集的外部验证。使用 26 个 SY 样本的相同预测集评估不同校准集的预测能力，并与基于偏最小二乘判别分析（PLS-DA）的传统一维（1D）NIR 光谱进行比较。结果表明，对于同一品牌模型，校准集和预测集的正确率分别为 100%和 96.15%。对于不同品牌模型，GM、MN 和 WDS 牛奶的 NPLS-DA 校准集正确率均为 100%。预测集的正确率分别为 73%、88.46%和 69.23%，明显高于 PLS-DA（分别为 50%、53.83%和 50%）。证明了模型鲁棒性对牛奶品牌的变化敏感。所提出的方法可以有效降低品牌对判别模型的影响。
