The at-line performance of two different NIRS instruments to predict major and minor cheese nutritional traits was evaluated. For this purpose, 158 samples from dairy products were collected and analysed by reference methods. Spectra were acquired using a transmittance and a reflectance instrument. Predictive equations were developed on the whole dataset or dividing samples in groups. Samples clustering was performed using pairwise Mahalanobis distance and centroid linkage algorithm. Prediction models for protein, fat, saturated fatty acids and minerals showed good prediction performances (R² > 0.80). Instrument configuration had a limited impact on prediction accuracy. Overall, clustering approach reduced prediction error but coefficient of determination also decreased. Prediction of minor compounds with models built from a large variety of cheeses could be useful for process control. Cluster approach is recommended for specific traits and cheese type, for the fine tuning of final product characteristics.

评估了两种不同 NIRS 仪器预测主要和次要奶酪营养特性的在线性能。为此，我们收集了 158 份乳制品样品，并通过参考方法进行了分析。使用透射率和反射率仪器获取光谱。预测方程是在整个数据集上开发的，或者是将样本分组。使用成对马氏距离和质心链接算法进行样本聚类。蛋白质、脂肪、饱和脂肪酸和矿物质的预测模型显示出良好的预测性能（R ² > 0.80）。仪器配置对预测精度的影响有限。总体而言，聚类方法降低了预测误差，但决定系数也降低了。使用由多种奶酪构建的模型来预测微量化合物可能有助于过程控制。建议针对特定特征和奶酪类型使用聚类方法，以微调最终产品特性。