为了提高乙醇固态发酵过程在线监测的精度, 开展了基于傅里叶近红外光谱(FT-NIRS)分析技术的乙醇固态发酵过程参数快速定量检测研究。采用联合区间偏最小二乘法(siPLS)对标准正态变量变换(SNV)预处理后的光谱进行特征波长区间优选; 引入遗传算法(GA)、竞争自适应重加权采样(CARS)法和迭代保留信息变量(IRIV)法从优选后波长区间中进一步筛选特征波长变量; 最后, 建立不同变量筛选方法所得特征波长的乙醇固态发酵过程参数(乙醇和还原糖含量)的偏最小二乘(PLS)预测模型。实验结果显示, 与GA和CARS方法相比, IRIV方法所得的波长变量数最少; 其中, 与乙醇和还原糖相关的特征变量个数分别为43和40; 在验证集中, PLS预测模型乙醇含量的验证集均方根误差(RMSEP)和预测相关系数Rp分别为0.2511和0.9934, 还原糖含量的RMSEP和Rp分别为0.1730和0.9926, 其预测精度亦高于其他方法所得结果。实验结果表明, 利用近红外光谱分析技术实现乙醇固态发酵过程关键参数的在线检测是可行的; 并且IRIV方法是一种有效近红外光谱特征波长优选方法, 可提高预测模型精度。

In order to improve the accuracy of on-line monitoring of ethanol solid-state fermentation process, We carried out the fast quantitative detection of ethanol solid state fermentation process parameters based on Fourier transform near infrared spectroscopy (FT-NIRS). The synergic interval partial least squares (siPLS) method was taken to select the optimal wavelength intervals from the standard normal variate transformation (SNV) preprocessing spectra. Characteristic wavelength variables were extracted from the optimal wavelength intervals by genetic algorithm (GA), competitive adaptive reweighted sampling (CARS) and iteratively retaining informative variables (IRIV). The partial least squares (PLS) forecast model of solid state fermentation process parameters (the level of ethanol and glucose) was established. The characteristic wavelength variables of these parameters were selected by different methods. The results show that compared with GA and CARS methods, IRIV method can select the fewest wavelength variables and the number of characteristic variables associated with ethanol and glucose is 43 and 40, respectively. In IRIV-siPLS model, root mean square error of prediction (RMSEP) is 0.2511 and correlation coefficient of prediction Rp is 0.9934 for ethanol, RMSEP is 0.1730 and Rp is 0.9926 for glucose in prediction set, and the prediction accuracy of the results is also higher than those of other methods. Based on the result, on-line detection of key process parameters in the process of solid-state fermentation is feasible with the near infrared spectral technology. IRIV is an effective method to select characteristic wavelength from the near infrared spectra and improve precision of the prediction model.