应用近红外光谱技术无损分析小麦种子短期自然老化过程中主要化学成分的变化趋势, 并结合支持向量机建立快速判别小麦种子自然老化程度的分析模型。 本实验应用VERTEX 70傅里叶变换红外光谱仪, 以大样品杯旋转采样方式跟踪采集了45份小麦种子在自然老化初期、 4个月、 7个月、 9个月的近红外光谱。 标准差可以用来表征数据离散程度, 因此本实验通过计算每份样本在4个自然老化阶段的光谱标准差来筛选与自然老化时间显著相关的谱区。 为避免单个样本由于偶然因素导致的离散度值异常, 实验统计了45份样本的光谱标准差均值, 根据均值光谱得到如下谱峰: 8 362, 6 950, 7 563, 5 319, 4 998和4 478 cm-1处。 解析谱峰所在区域对应的化学基团归属可得: 6 950 cm-1处对应的是液态水中O—H伸缩振动的一级倍频且该处离散度值较大, 因此小麦种子在短期自然老化阶段中水分变化较为显著; 5 319, 4 998和4 478 cm-1处离散度值较6 950 cm-1处小, 对应的是蛋白质仲酰胺、 伯酰胺和酰胺的合频和倍频信息, 因此蛋白质变化较水分而言相对平缓; 8 362和7 563 cm-1处反映的主要是C—H振动的二级倍频信息且离散度值较大, 而种子中蛋白质、 淀粉等均具有C—H官能团, 因此蛋白和淀粉等成分综合变化较为显著。 在上述分析基础上, 本文采用多分类支持向量机结合近红外光谱建立快速识别小麦种子四种自然老化程度的定性模型。 将180份样本光谱按照3∶1随机抽取135个样本作为训练集, 其余样本作为测试集。 选择核函数为径向基函数, 通过网格搜索法进行参数寻优得到惩罚参数为8, 核参数为0.008 974 2时, 训练集和测试集的识别正确率可达99.26％和99.78％。 实验结果表明: 近红外光谱技术结合支持向量机可快速判别小麦种子短期自然老化程度, 为种子贮藏过程中生理特性变化的无损监测及开发利用提供便捷的检测手段。

To study the variation trend of major chemical composition of wheat seeds during short-time natural aging, the nondestructive technology based on near infrared spectroscopy (NIR) and support vector machines (SVM) is applied to evaluate the natural aging stage at the same time. There are 45 wheat samples collected in the experiment. The samples are scanned at the beginning and after natural aging for 4 months, 7 months and 9 months respectively by VERTEX 70 Fourier transform infrared spectrometer in large sample cup rotation sampling mode. The spectral standard deviations of each sample at four natural aging stages are calculated firstly. The standard deviations represent the statistical quantity of data dispersion. The obvious variation regions are screened according to the standard deviations calculated from the spectrums of 4 aging stages. To avoid abnormal discrete degree value caused by accidental factors, the averages of 45 samples spectrum discrete degree are calculated. The spectral peaks are mainly distributed in the area of 8 362, 6 950, 7 563, 5 319, 4 998 and 4 478 cm-1 according to the standard deviation. The region nearby 6 950 cm-1 reflects stretching vibration of O—H in liquid water, and the standard deviation value is greater. This illustrates the moisture changes remarkably during natural aging stage. The region nearby 5 319, 4 998 and 4 478 cm-1 reflect vibration information of primary amide, secondary amide and amide in protein. The standard deviation values at these peaks are all lower than the value of 6 950 cm-1, so the protein changes more slowly than moisture during aging stage. The region nearby 8 362 and 7 563cm-1 reflect secondary vibration information of C—H and the he standard deviation value is greater. There are C—H group in protein, starch, etc. of wheat seeds. It shows that comprehensive changes of protein, starch and other components are relatively strong. According to the above analysis, the multi-classification model has been built based on NIR and SVM to determine the 4 types natural aging stages. The sample set is divided into two parts randomly according to the ratio of 3∶1. The number of train sample is 135 and the number of test sample is 45. The best parameters of SVM are selected by grid searching. While the kernel function is RBF function, the penalty parameter is 8 and kernel parameter is 0.008 974 2, and the recognition rate of training set and test set reach to 99.26% and 99.78%. The results show that NIR technology combined with SVM can be applied to determine the natural aging stage of wheat seeds, which also provides a convenient and fast tool to monitor physiological characteristics changes during wheat seeds storage.