双光谱区间遗传算法及其在模型转移中的应用

在近红外光谱分析中，将近红外光谱和浓度信息建立统计模型，通过光谱代入模型即可预测未知样本浓度。但是，检测条件的变化会导致光谱的改变，进而导致原有的模型不能准确预测光谱改变后的样本。对此，模型转移可以通过校正新测量的光谱（从光谱），使得从光谱能够被原有光谱（主光谱）建立的模型准确预测。模型转移可以使用全光谱进行校正，但是全光谱中往往包括噪声、背景等干扰信息，这些干扰会增加预测误差。故可以使用变量选择方法找出光谱中有化学意义的信息来模型转移。但是一般的变量选择算法只选择主光谱的区间，从光谱使用主光谱相同的波长区间模型转移。但是在实际工作中，主光谱和从光谱有化学意义的区间往往不一致，主从光谱使用同一区间模型转移会增加误差；此外，有时二者原光谱的波长范围并不一致，从主光谱选出的区间不能用于从光谱的校正。对此，提出了基于双光谱区间遗传算法（GA-IDS），同时选择主光谱和从光谱有化学意义的区间，进而实现模型转移。GA-IDS算法步骤包括，①随机产生种群；②分析种群中每条染色体，删去错误染色体；③根据每条染色体，找出其相应的主光谱和从光谱波段组合，并计算其模型转移后的验证均方根误差(RMSEV)；④按照概率，执行选择、交叉、变异操作。在一次迭代结束之后，返回到步骤②，重新执行纠错、计算RMSEV、选择、交叉、变异。达到停止迭代的要求后，将最低的RMSEV值所对应的染色体保存下来作为最优染色体，其所对应的主从光谱区间作为最优区间。用玉米、小麦两套数据测试了该算法，结果显示，与全光谱相比，GA-IDS选择的主从光谱区间可以显著地降低误差；与向后迭代区间选择法（IIBS）相比，在小样本情况下，GA-IDS的误差显著地小于IIBS方法。

Interval Genetic Algorithm for Double Spectra and Its Applications in Calibration Transfer

As a non-destructive detection method, near-infrared (NIR) spectra have been widely used in food analysis. In NIR analysis, the model between spectra and sample concentrations should be calibrated in advance, and the concentrations of new samples can be predicted by substituting their spectra with the calibrated model. However, the variation of measurement conditions can lead to spectra changes. This problem can be solved by calibration transfer which corrects the new spectra (secondary spectra) to be accurately predicted by the old spectra (primary spectra) model. The calibration transfer always uses full primary and secondary spectra for correction. However, full primary and secondary spectra contain interference, including noise and background, which can increase prediction errors. Hence, variable select is used to selecting the informative regions of NIR for calibration transfer. The commonly used variable selection method always treats primary spectra, and both primary and secondary spectra share the same regions for calibration transfer. However, in practical work, the informative regions of primary and secondary spectra are not the same. Thus, both primary and secondary spectra using the same regions for calibration transfer can increase prediction errors. Moreover, the original spectral ranges of primary and secondary spectra may not be the same, and the secondary spectra can not use the regions selected by primary spectra for calibration transfer. In order to solve this problem, this paper proposed a Genetic algorithm for intervals of double spectra (GA-IDS), which selects informative regions for both primary and secondary spectra simultaneously for calibration transfer. The procedure of GA-IDS includes(1)Randomly generating chromosomes in the population;(2)Analyzing each one of the chromosomes and deleting the error ones; (3)Obtaining the primary and secondary spectra regions and the corresponding Root mean squared error of validation (RMSEV) based on each one of the chromosomes; (4)Executing selection, crossover and mutation operations. After finishing one loop, the GA-IDS goes to step (2) to repeat execute errors correction, RMSEV computation, selection, crossover and mutation operation. After achieving the criterion of the final termination, the spectra regions with minimal RMSEV can be retained. Two datasets, including corn and wheat datasets, were used to evaluate this algorithm. The results show that, compared with full variables, GA-IDS can select good regions for both primary and secondary spectra to reduce prediction errors. Compared with Iterative interval backward selection (IIBS), GA-IDS can achieve lower errors at the small size transfer set.