便携式近红外光谱与集成学习法快速鉴别8种食用淀粉

Rapid identification of 8 kinds of edible starch species by portable near infrared spectroscopy and ensemble learning

中文关键词:  近红外光谱  淀粉物种  机器学习  快速检测  便携式设备

英文关键词:near infrared spectroscopy  starch species  ensemble learning  rapid detection  portable device

中文摘要:

      目的 构建基于便携式近红外光谱(1000~1800 nm)与集成学习算法的智能检测系统快速鉴别8种食用淀粉。方法 采集了玉米、马铃薯、木薯、小麦、红薯、绿豆、藕、豌豆等8类共239份市售淀粉样品, 获取2392条近红外光谱。采用标准正态变量变换(standard normal variate, SNV)和Savitzky-Golay一阶导数(窗口长度11, 多项式阶数2)进行预处理, 并直接在全谱范围内建模, 未进行特征波长筛选。结果 训练并对比了K近邻算法(K-nearest neighbor, KNN)、随机森林算法(random forest, RF)、偏最小二乘判别分析(partial least squares discriminant analysis, PLS-DA)、极端梯度提升库(eXtreme gradient boosting, XGBoost)及集成模型5种分类模型。在59份独立样品的验证中, 集成模型的准确率达97.0%, 布里尔分数低至0.008。结论 该系统单次预测耗时仅0.8 s, 为淀粉掺假监测和物种溯源提供了一种高效、可靠的现场检测技术方案。

英文摘要:

      Objective To construct an intelligent detection system based on portable near infrared spectroscopy (1000–1800 nm) and ensemble learning algorithms for rapid identification of 8 kinds of edible starch. Methods A total of 239 commercial starch samples from 8 botanical sources—corn, potato, cassava, wheat, sweet potato, mung bean, lotus root and pea—were collected, yielding 2392 near infrared spectroscopy spectra. The spectra were preprocessed by standard normal variate (SNV) and Savitzky-Golay first derivative (window length=11, polynomial order=2). And it was directly modeled across the full spectral range without conducting characteristic wavelength screening. Results The 5 kinds of classification models—K-nearest neighbors (KNN), random forest (RF), partial least squares discriminant analysis (PLS-DA), eXtreme gradient boosting (XGBoost), and a weighted soft-voting ensemble model—were trained and compared. In validation using an independent test set of 59 samples, the ensemble model achieved an accuracy of 97.0% with a Brier score as low as 0.008. Conclusion The system takes only 0.8 seconds for a single prediction, providing an efficient and reliable on-site detection technical solution for starch adulteration monitoring and species traceability.