近年来, 深度学习在数据挖掘领域研究较多, 深度学习中的集成学习算法也越来越多地应用到分类和定量回归中, 但是, 集成学习算法在红外光谱分析领域的应用研究较少。 提出一种基于Blending模型融合的集成学习定量回归算法, 利用GBDT算法、 线性核支持向量机(LinearSVM)和径向基核支持向量机(RBF SVM)作为基学习器, 将基学习器预测结果通过LinearSVM模型完成数据融合。 以公开数据库中的药片和柴油近红外光谱数据为研究对象, 首先对光谱数据进行一阶导数预处理, 分别采用单核支持向量回归模型、 GBDT模型和Blending集成学习模型, 将模型预测结果进行分析比较。 药片活性物含量和硬度性质采用RBF SVM模型的预测结果最优, RMSEP最小, RPD最大; 其次为Blending集成学习模型; GBDT模型预测结果最差。 药片质量采用Blending集成学习模型预测的 R²最高, 达到0.837 4; RBF SVM的RMSEP最小, 为2.140 6, RPD最大, 达到7.487 8; LinearSVM的预测结果最差。 对于柴油沸点、 闪点和总芳香烃三种性质, Blending模型预测效果最好, 优于三种单模型预测结果。 对于十六烷值, GBDT模型和RBF SVM模型预测结果优于Blending集成学习模型。 对于密度, 仅GBDT模型优于Blending集成模型, 并且, 使用单模型和集成模型的预测结果均较为理想, 除了LinearSVM模型 R²为0.944 5, 其他模型 R²均高于0.99。 对于冰点的预测, RBF SVM和LinearSVM的预测效果优于Blending集成学习模型。 对于黏性性质的预测, 仅RBF SVM的预测效果优于Blending集成算法模型。 由结果可以看出, 由GBDT, LinearSVM和RBF SVM集成的Blending模型由于融合了单模型的特征, 与单模型相比, 预测效果较优或者最优, 证明集成学习Blending模型用于红外光谱定量回归具有较强的适用性, 且具有较高的预测精度和泛化能力, 对于进一步研究集成学习算法在红外光谱定量回归中的应用具有重要的意义。

In recent years, deep learning has been studied more and more in the field of data mining, and the integrated learning algorithm in deep learning has been applied to classification and quantitative regression more and more, but the application of integrated learning in the field of infrared spectrum analysis is little. In this paper, an integrated learning quantitative regression algorithm based on Blending model is proposed. GBDT algorithm, linear kernel support vector machine (LinearSVM) and radial kernel support vector machine (RBF SVM) are used as the basic learners, and the prediction results of the basic learners are fused by LinearSVM. The first derivative preprocessing was carried out for the spectral data. The prediction results of the model were analyzed and compared by using the GBDT, LinearSVM, RBF SVM and the Blending integrated learning model respectively. RBF SVM model is the best model for predicting the content of active substance and hardness, R² is the highest, the RMSEP is the smallest, and the RPD is the largest, and the GBDT model is the worst. The R² of tablet quality predicted by Blending model is the highest, reaching 0.837 4, while the RMSEP of RBF SVM is the lowest, 2.140 6, and the RPD of RBF SVM, 7.487 8, is the largest. For the boiling point, flash point and total aromatics of diesel oil, Blending model is the best one, which is better than the single model. For the cetane number, GBDT model and RBF SVM model are better than Blending model. For the density property, the single model and the integrated model have better prediction results, except that the R² of LinearSVM model is 0.944 5, R² of other models are all higher than 0.99. For the prediction of freezing point properties, RBF SVM and LinearSVM are both better than Blending model. For the prediction of viscosity, only RBF SVM is better than Blending model. It can be seen from the results that the Blending model integrates the characteristics of GBDT, LinearSVM and RBF SVM model, compared with the single model, the prediction of Blending is better or optimal. It is proved that Blending integrated learning model has strong applicability for infrared quantitative regression, and has a high prediction accuracy and generalization ability. It is of great significance for further research on the application of integrated learning algorithm in infrared quantitative regression.