基于移动窗口-迭代遗传算法的近红外光谱波长选择方法  

Near Infrared Spectral Wavelength Selection Based on Moving Window-Iterative Genetic Algorithm Method

作　　者：成飙[1] 陈德钊[1] 吴晓华[1] 

机构地区：[1]浙江大学化学工程与生物工程学系,杭州310027

出　　处：《分析化学》2006年第U09期123-126,共4页Chinese Journal of Analytical Chemistry

基　　金：浙江省重点科技项目基金资助(No.2004C21SA120002)

摘　　要：光谱样本数据常会受到环境噪声和其它组分的干扰,应作波长选择,以提高分析精度。近红外光谱谱区宽,搜索空间过大,难以直接采用遗传算法进行波长选择。为此本研究提出先用移动窗口偏最小二乘法(MWPLS)从宽谱区中初选出信息区间,再采用改进的迭代遗传算法(IGA)从中选出最优的信息子区间。MWPLS用移动窗口沿全谱区扫描,对信息区间的定位效果好,而IGA将顾及光谱数据的连续相关特性,运行多轮GA,并以上轮选择结果平滑处理后作为先验知识支持下轮的种群初始化。由此选出的连续相邻的波长点作为自变量,进行PLS建模,既可显著地简化模型,又保留一定的数据冗余,模型的稳健性好,分析精度高。将其用于小麦水分的近红外分析,效果良好,预测性能明显优于其它方法。Standard genetic algorithm is usually not fit to near infrared spectral（NIR） wavelength selection, because the number of NIR spectral wavelengths is too large. Further more, standard genetic algorithm（GA） can select a few wavelengths only, so the model can not exploit the advantage of partial least square （PLS）. A two-stages strategy was proposed. Firstly, along the whole spectral region, a series of PLS models were built in moving window, and then the informative regions with small error could be located. Secondly, iterative genetic algorithm（IGA） was used to select sub-regions in the informative regions. IGA gave an initializing procedure of selecting the first candidates for every run of GA, which uses the results of previous runs as guiding information. This algorithm can select wavelength regions instead of scattering points, and the collinear data with some redundancies make the PLS model more robust. But the selected wavelengths are not too many, so the model is simplified. The result obtained on wheat dataset demonstrates the two stages strategy is an effective and robust NIR wavelength selection approach.

关 键 词：近红外光谱 波长选择 移动窗口 偏最小二乘 遗传算法 

分 类 号：O657.33[理学—分析化学]