局部偏最小二乘法结合可见-近红外光谱预测猪肉挥发性盐基氮

Prediction of Total Volatile Basic Nitrogen in Pork Using Local Partial Least Squares Combined with Visible and Near-Infrared Spectroscopy

以两个批次的猪肉为实验样品,开展局部偏最小二乘法结合双波段可见-近红外光谱预测挥发性盐基氮的研究,以改善模型预测不同批次样品时效果不佳的问题。提出基于距离、信息测度和投影的相似性度量方法,通过对欧式距离和光谱信息散度-光谱角(spectral information divergence-spectral angle,SID-SAM)进行加权求和,构建评价不同样品相似性的相似度函数,定义相似度因子(SM),通过最小化代价函数确定建立局部模型的邻域窗口。以第1批样品为建模基础集,通过对欧式距离和SID-SAM的权重及SM进行参数寻优,针对第2批次中每个样品建立局部偏最小二乘模型。结果显示,与利用第1批样品建立的模型直接预测第2批样品的结果相比,预测效果有明显提高,相关系数R从0.845 6上升至0.948 1,预测误差从4.581 0 mg/100 g下降至2.650 8 mg/100 g。这表明利用提出的相似度函数和相似度因子,可根据待测样品的光谱特征,实时动态选择相似的局部空间,建立的局部偏最小二乘模型能有效提高对外部验证样品的预测能力。

In this study, two batches of pork were used as experimental samples, and local partial least squares(LPLS) and dual-band visible and near-infrared spectroscopy were used in conjunction to build a prediction model for the detection of total volatile basic nitrogen(TVB-N). A similarity measurement method based on distance, information measurement, and projection was proposed. A similarity function to evaluate the similarity of different samples was also proposed by weighted summation of Euclidean distance and spectral information divergence-spectral angle(SID-SAM). Then, a similarity factor(SM) was defined and employed to determine the modeling neighborhood window for the establishment of LPLS model.Using thefirst batch of samples as the modeling basis set, LPLS models for each sample in the second batch of samples were built by optimizing the weights of Euclidean distance and SID-SAM and SM. Compared with the results obtained before establishing the LPLS models, the correlation coefficient(R) for the second batch of samples increased from 0.845 6 to 0.948 1, and the prediction error decreased from 4.581 0 to 2.650 8 mg/100 g, which improved the prediction accuracy of the model. The results showed that using the proposed similarity function and similarity factor, a local space can be dynamically selected in real time according to the spectral characteristics of the samples. The established LPLS prediction models can effectively improve the prediction ability for the validated samples.