针对近红外光谱定性分析中, 增加新的品种进行建模时, 原有模型识别效果不够稳定的问题, 提出一种在建模样本的基础上添加同类物质的历史光谱数据的特征提取方法, 首先采集建模样本的近红外光谱数据, 然后添加同种物质样本的历史近红外光谱数据, 再对所有近红外光谱数据进行预处理, 其次对所有样本数据进行偏最小二乘(PLS)特征提取得到偏最小二乘空间, 并只将建模样本数据向构建的偏最小二乘空间进行投影, 最后将投影后的建模数据进行正交线性判别分析(OLDA)特征提取。 以玉米种子近红外光谱为研究对象, 分别对建模数据添加历史近红外光谱以及不添加历史近红外光谱两种情况进行特征提取, 并通过仿生模式识别(BPR)方法构建模型进行验证, 实验结果表明, 添加历史近红外光谱构建偏最小二乘空间的特征提取方法相对于不添加历史近红外光谱的方法, 首先在增加建模集品种数量时, 原有的品种识别率基本不变; 其次在相同PLS维数时, 所建模型对不同时间采集的测试集识别效果基本一致, 证明了该方法可以提高模型稳健性。 在实际应用中就可以在品种鉴别软件中将特征提取维数设置为固定值, 免除了品种鉴别软件的用户在增加建模集品种时为了保证最优识别效果重新选定最优PLS参数的麻烦。

In traditional qualitative analysis of near-infrared (NIR) spectra, the stability of recognition models is decreased when new varieties of samples are added into the model. In order to improve the robustness of the model, a new feature extraction method based on the addition of historical data was put forward. The NIR training samples will be collected first, after that the historical data of the same species is added to constitute a larger and richer dataset. Then, the pretreated data of these training samples is projected to the feature space, which is constructed by feature extraction using partial least squares (PLS) based on the above dataset. Subsequently, orthogonal linear discriminant analysis (OLDA) is employed to extract features of the projected data. 18 varieties of corn seeds were taken as study subject, the comparative experiments with and without historical data are implemented respectively, and then the biomimetic pattern recognition (BPR) method is applied to verify the efficiency of the method proposed. The results suggest that the method adopted can improve the robustness of recognition model more effectively compared with the method without historical data. It maintains the high correct recognition ratios when new varieties are added into the model. Besides that, the recognition effect on test sets of the different days remains the same basically in the condition of same PLS dimensions. Therefore, the dimension of feature extraction can be set to some fixed values in recognition software. In this way, it can keep out of the trouble of manually modifying the optimal PLS parameter in recognition software if new varieties need to be added into the model. The experiment results of the thesis manifested the effectiveness of the proposed method.