【摘要】 木糖作为一种功能性低聚糖,具有抗氧化、促进肠道健康等保健作用,被广泛应用于食品、医药和生物燃料等领域。目前仍然缺乏有效的木糖含量快速检测方法。针对木糖生产过程中的含量检测问题,提出一种基于近红外光谱的在线检测方法。首先,采集样本溶液,使用近红外光谱仪扫描得到原始光谱,进一步使用一阶导数和平滑滤波方法对原始光谱进行预处理,去除噪声与基线漂移的影响。然后,使用随机蛙跳算法对光谱变量进行特征选择,结合预测相对分析误差搜索最佳特征数,结果显示:当特征数在20～30时模型的预测性能最优。综合其他指标,选择特征数量为25,确定能表征木糖含量的波长特征。由于随机蛙跳算法具有随机子集选择和随机森林回归的特性,该算法在执行高维木糖数据的特征波长筛选任务时存在明显优势,同时也存在结果重现性低的缺陷。在得到波长特征后对结果进行加权累计,弱化该算法不确定性对最终模型的影响,再以液相色谱仪测得的数据为标签,建立木糖含量的预测模型。最后,使用以上方法对工艺现场采集的样本进行木糖含量的快速测定,并对比PLS模型及Lasso模型的预测效果。结果表明,指标中的训练集决定系数R~2=0.9377,测试集决定系数R_p~2=0.9335,R~2和R_p~2都接近1,模型能较好的解释训练集数据,并具备良好的泛化性能。预测均方根误差RMSEP=5.8446,预测相对分析误差RPD=3.8792>2.5,模型可以较为准确地预测样品的木糖含量。通过对比发现,RJFA-PLS模型的各项评价指标均优于PLS模型,RMSEP降低了112.7%,R~2、RPD和R_p~2分别提高了21.8%、52.5%和24.6%。而Lasso算法在基于本数据集的木糖含量预测上表现不佳。在此次实验条件下,使用以上方法建立的模型比PLS模型及Lasso模型更适用于木糖含量的预测。本方法解决了木糖含量检测结果滞后的问题,还为木糖在线检测技术的研究提供了先决条件。更多还原

【Abstract】 Xylose,as a functional oligosaccharide,possesses health benefits such as antioxidant properties and promoting intestinal health,and is widely used in food,medicine,and biofuels.There is still a lack of effective rapid detection methods for xylose content.An online detection method based on near-infrared spectroscopy technology is proposed to address the issue of content detection during xylose production.Firstly,sample solutions are collected and scanned using a near-infrared spectrometer to obtain raw spectra.The raw spectra are then preprocessed using first derivative and smoothing filter methods to remove noise and baseline drift effects.Subsequently,the random frog algorithm is employed for feature selection of spectral variables,and the prediction relative analysis error is used to search for the optimal number of features.The results show that the model’s predictive performance is optimal when the number of features is between 20 and 30.Considering other indicators,the number of features is selected as 25,determining the wavelength characteristics representing xylose content.Due to the random subset selection and random forest regression characteristics of the random frog algorithm,this algorithm has obvious advantages in performing the task of feature wavelength screening for high-dimensional xylose data,but also has the defect of low result reproducibility.After obtaining the wavelength features,the results are weighted and accumulated to weaken the impact of the algorithm’ s uncertainty on the final model.Then,a predictive model for xylose content is established using data measured by a liquid chromatograph as labels.Finally,the method is used to rapidly determine the xylose content of samples collected from the process site,and the prediction effects are compared with those of the PLS and Lasso models.The results indicate that the training set determination coefficient R2=0.937 7,and the test set determination coefficient R_p~2=0.933 5,with R~2 and R_p~2 close to 1,indicating that the model can explain the training set data well and has good generalization performance.The prediction root mean square error RMSEP=5.844 6,and the prediction relative analysis error RPD=3.879 2>2.5,indicating that the model can predict the xylose content of samples relatively accurately.Through comparison,it is found that the RJFA-PLS model’s evaluation indicators are superior to those of the PLS model,with RMSEP reduced by 112.7%,and R~2,RPD,and R_p~2 increased by 21.8%,52.5%,and 24.6%,respectively.However,the Lasso algorithm performs poorly predicting xylose content based on this dataset.Under the experimental conditions of this study,the model established using the above method is more suitable for predicting xylose content than the PLS and Lasso models.The proposal of this method solves the problem of lag in xylose content detection results and also provides a prerequisite for the research of online detection technology for xylose.更多还原