在近红外光谱分析中, 向前间隔偏最小二乘法（FiPLS）和向后间隔偏最小二乘法（BiPLS）是常用的基于波长变量选择的建模方法, 其模型精度较高, 但贪婪搜索特性较强, 导致选出的波段并不能较好地反映待测成分的信息。 针对该问题, 提出一种基于两者组合策略的光谱特征波段选择方法（FB-iPLS）。 在光谱分段的基础上, 既利用FiPLS选取有用波段, 同时利用BiPLS删除无用波段, 来交互执行特征变量的选择与删除, 对目标特征波段进行双向选择, 用于提高模型的稳健性。 用该方法建立水中乙醇含量的定量预测模型, 并与FiPLS和BiPLS算法对比。 由于光谱分段大小会对模型的结果有影响, 该实验还考查这三种方法在不同光谱分段处的结果。 在光谱划分60段时, 提出的FB-iPLS方法取得最佳预测性能, 其校正集与验证集相关系数r分别为0.967 7, 0.967 0, 交互验证均方根误差RMSECV分别为0.088 8, 0.057 1。 与FiPLS和BiPLS相比, 该方法无论在不同光谱分段区间还是在各自最优与最差分段处, 模型的整体预测性能都有所提高。 实验结果表明, 提出的方法能改善BiPLS与FiPLS贪婪搜索的特性, 对特征波段的选取更高效、 更具代表性, 能进一步提高模型的预测性能。

In the near-infrared spectroscopy, the Forward Interval Partial Least Squares (FiPLS) and Backward Interval Partial Least Squares (BiPLS) are commonly used modeling methods, which are based on the wavelength variable selection. These methods are usually of high prediction accuracy, but are strongly characteristic of greedy search, which causes that the intervals selected are not good enough to indicate the analyte information. To solve the problem, a spectral characteristic intervals selection strategy (FB-iPLS) based on the combination of FiPLS and BiPLS is proposed. On the basis of spectral segmentation, both FiPLSs are used to select useful intervals, and BiPLS is used to delete useless intervals, so as to perform the selection and deletion of the characteristic variables alternatively, which conducts a two-way choice of the target characteristic variables, and is used to improve the robustness of the model. The experiments on determining the ethanol concentration in pure water are conducted by modeling with FiPLS, BiPLS and the proposed method. Since different size of intervals will affect the result of the model, the experiments here will also examine the model results with different intervals of these three models. When the spectrum is divided into 60 segments, the FB-iPLS method obtains the best prediction performance. The correlation coefficients (r) of the calibration set and validation set are 0.967 7 and 0.967 0 respectively, and the cross-validation root mean square errors (RMSECV) are 0.088 8 and 0.057 1, respectively. Compared with FiPLS and BiPLS, the overall prediction performance of the proposed model is better. The experiments show that the proposed method can further improve the predictive performance of the model by resolving the greedy search feature against BiPLS and FiPLS, which is more efficient for and representative of the selection of characteristic intervals.