为了实现固态发酵过程状态的快速监测, 以饲料蛋白固态发酵为实验对象, 开展了基于近红外光谱分析技术的饲料蛋白固态发酵过程状态定性识别研究。 首先利用Antaris Ⅱ型傅里叶变换近红外光谱仪采集140个固态发酵物样本的近红外光谱, 并采用标准正态变换(SNV)光谱预处理方法对获得的原始光谱进行预处理; 其次, 采用谱回归判别分析(SRDA)法对预处理后的近红外光谱进行特征提取; 最后, 采用最近邻(NN)分类算法作为弱分类器建立固态发酵过程状态识别模型, 并对测试集样本进行识别。 结果显示, 与利用主成分分析(PCA)法和线性判别分析(LDA)法提取的光谱特征建立的识别模型结果相比较, SRDA-NN识别模型获得的结果最佳, 在测试集中的正确识别率达到94.28%; 为了进一步提高识别模型的准确率, 将自适应提升法(Adaboost)与SRDA-NN方法结合, 提出了Adaboost-SRDA-NN集成学习算法来建立饲料蛋白固态发酵过程状态的在线监测模型。 通过Adaboost算法提升后的SRDA-NN模型预测性能得到了进一步增强, Adaboost-SRDA-NN模型在测试集中的正确识别率达到100%。 试验结果表明: 在近红外光谱定性分析模型校正过程中, SRDA方法能有效地对近红外光谱数据进行特征提取, 以实现维数约简; 另外, Adaboost算法能很好地提升最终分类模型的预测精度。

In order to achieve the rapid monitoring of process state of solid state fermentation (SSF), this study attempted to qualitative identification of process state of SSF of feed protein by use of Fourier transform near infrared (FT-NIR) spectroscopy analysis technique. Even more specifically, the FT-NIR spectroscopy combined with Adaboost-SRDA-NN integrated learning algorithm as an ideal analysis tool was used to accurately and rapidly monitor chemical and physical changes in SSF of feed protein without the need for chemical analysis. Firstly, the raw spectra of all the 140 fermentation samples obtained were collected by use of Fourier transform near infrared spectrometer (Antaris Ⅱ), and the raw spectra obtained were preprocessed by use of standard normal variate transformation (SNV) spectral preprocessing algorithm. Thereafter, the characteristic information of the preprocessed spectra was extracted by use of spectral regression discriminant analysis (SRDA). Finally, nearest neighbors (NN) algorithm as a basic classifier was selected and building state recognition model to identify different fermentation samples in the validation set. Experimental results showed as follows: the SRDA-NN model revealed its superior performance by compared with other two different NN models, which were developed by use of the feature information form principal component analysis (PCA) and linear discriminant analysis (LDA), and the correct recognition rate of SRDA-NN model achieved 94.28% in the validation set. In this work, in order to further improve the recognition accuracy of the final model, Adaboost-SRDA-NN ensemble learning algorithm was proposed by integrated the Adaboost and SRDA-NN methods, and the presented algorithm was used to construct the online monitoring model of process state of SSF of feed protein. Experimental results showed as follows: the prediction performance of SRDA-NN model has been further enhanced by use of Adaboost lifting algorithm, and the correct recognition rate of the Adaboost-SRDA-NN model achieved 100% in the validation set. The overall results demonstrate that SRDA algorithm can effectively achieve the spectral feature information extraction to the spectral dimension reduction in model calibration process of qualitative analysis of NIR spectroscopy. In addition, the Adaboost lifting algorithm can improve the classification accuracy of the final model. The results obtained in this work can provide research foundation for developing online monitoring instruments for the monitoring of SSF process.