目的研究偏最小二乘-判别分析模型的影响因素。方法以小麦为例,采用人工加速老化结合近、中红外光谱技术采集陈化、非陈化小麦粉的近、中红外光谱,分别针对小麦粉衰减全反射中红外数据、衰减全反射中红外一阶导数数据、近红外一阶导数数据以及小麦颗粒近红外一阶导数数据建立偏最小二乘-判别分析模型,分别获取模型参数及第1、2主成分得分分布情况。结果小麦粉近红外一阶导数建模的第1、2主成分得分散点图具有明显的两类样本分布。外部验证数据表明,小麦粉近红外一阶导数所建模型具有更好的稳健性。结论样品的物理形态、数据预处理以及红外光谱波段皆会对偏最小二乘-判别分析模型结果产生影响;同时,红外光谱偏最小二乘-判别分析模型的评价需要从正确识别率、模型的校正测定系数、交互验证测定系数、校正均方根误差、交互验证均方根误差以及模型主要成分得分分布等多种情况综合考虑。 Objective To study the influencing factors of partial least squares discriminant analysis model. Methods Taking wheat as an example, the near and mid-infrared spectra of aged and un-aged wheat flour were collected by artificial accelerated aging combined with near and mid-infrared spectroscopy. The mid-infrared data of attenuated total reflection and the first derivative of attenuated total reflection Partial least squares discriminant analysis model was established based on the first derivative data of near infrared and the first derivative data of near infrared and the first and second principal component scores were obtained respectively. Results The first and second principal components scoring scatter plots of the first derivative of wheat near-infrared have obvious two types of sample distribution. External validation data show that the wheat near infrared first derivative model has better robustness. Conclusion The physical shape of the sample, data preprocessing and the band of infrared spectrum all affect the results of partial least squares discriminant analysis model. Meanwhile, the estimation of partial least squares discriminant analysis model of infrared spectrum needs to be based on the correct recognition rate, Correction coefficient, cross-validation coefficient, root mean square error correction, cross validation root mean square error and the main components of the model score distribution and other conditions considered.