针对近红外光谱数据的特点, 在分析了单模型波长选择方法的基础上, 提出了一种多模型融合的变量选择方法。 它融合多个模型的回归系数, 以提高波长选择的准确性和稳定性。 并用3个业界标准的近红外光谱数据集对提出的方法进行了验证, 同时与UVE-PLS和GA-PLS算法进行了比较。 实验结果表明, 经该方法选择变量后, 提高了模型的预测能力, 降低了复杂度, 达到甚至优于UVE-PLS和GA-PLS, 而且具有算法简单、 效率高的优点, 具有广泛的实用价值

NIR spectroscopy makes a feature of a large number of wavelengths with a much smaller set of samples. However, some of the wavelengths contribute no information to the modeling. Even worse, they may contain the irrelevant information such as noise and background, which may result in a complex model and/or bad predictive ability of the model. So, it’s important to do research in-depth to eliminate these wavelengths and improve the quality of the final model. The present paper firstly summarizes the variable selection methods based on a single PLS regression model and concludes that (1) the cross-validation can be used to select optimal model with good predictive ability, but the resulting model may be not suitable for selecting variables; (2) selecting variables based on a single regression model is inaccurate and instable because a single vector of regression coefficients may not measure the importance of the variables correctly and may vary with models of different complexity. On basis of this analysis, this paper proposed a new method for variable selection based on the fusion of multiple PLS models. This method fuses the multiple PLS regression coefficients to form a vector, then a threshold is determined to eliminate the variables whose corresponding element in the vector is lower than this threshold. Finally, this method is verified by 3 well-known NIR datasets and compared with the UVE-PLS and GA-PLS algorithms. The experiments show that this method may result in a model with less complexity and/or better predictive ability. Moreover, the proposed method is elegant and efficient and therefore can be put in practical use.