针对油脂脱臭过程中的反式脂肪酸(TFAs)含量控制问题, 提出一种基于近红外光谱分析的油脂中TFAs含量快速检测方法。 制备含不同TFAs的大豆油脂样本100个, 利用气相色谱(GC)法精确测定其TFAs含量, 扫描样本近红外光谱, 然后利用不同方法对光谱数据进行降噪处理, 发现多元散射校正的去噪效果最佳。 为了探讨TFAs在近红外区域的吸收特性, 采用多种iPLS方法对比分析, 筛选出7 258～7 443/6 502～6 691/6 120～6 309 cm-1 TFAs的特征波段, 再利用Kalman滤波算法进行特征波长变量的选择, 优选出27个TFAs的特征波长变量; 采用深度信念网络(DBN)建立校正模型, 通过多次对比发现, 当隐含层层数为3并且隐含层节点数为50-35-90时, DBN模型性能最佳。 最后将DBN模型与PLS方法建立的反式脂肪酸含量回归模型进行对比分析, 结果表明: 对降噪后的全谱进行建模, DBN模型的预测效果优于PLS, DBN模型预测集R2为0.879 4、 RMSEP为0.060 3、 RSD为2.18％; 对筛选出的特征波段建模, PLS模型的预测效果优于DBN模型; 对优选出来的27个特征波长变量建模, DBN的预测效果较好, R2为0.958 4、 RMSEP为0.035 0、 RSD为1.31％, 说明DBN模型的泛化能力更好, 并且利用少量的波长变量就能达到较好的预测效果, 能够满足实际检测需求, 为实现油脂加工过程中TFAs含量的在线检测和调控, 生产低/零TFAs油脂产品提供技术支撑。

In order to control the content of trans fatty acids (TFAs) in the process of oil deodorization, this paper presents a fast method for detecting trans fatty acids (TFAs) content in soybean oil based on near-infrared spectroscopy. First, we prepared 100 soybean oil samples with different TFAs content, and detected precisely the values of TFAs contents by gas chromatography. Then, the near-infrared spectrum of oil samples was scanned and denoised by various methods, and it is found that the denoising effect of MSC was the best. In order to study the characteristic absorption of TFAs in near-infrared region, we used a variety of iPLS methods to select the characteristic band of the spectral data, and the characteristic absorption band of TFAs is selected as 7 258～7 443/6 502～6 691/6 120～6 309 cm-1. On this basis, the Kalman filtering algorithm is used to select the characteristic wavelength variables, and 27 TFAs characteristic wavelength variables are optimized. The deep belief network (DBN) is adopted to construct the correction model, and we found that the performance of the DBN model is the best adopting 3 hidden layers and 50-35-90 hidden layer nodes. Finally, the DBN model with this parameter is compared with the regression model of trans fatty acid content established by PLS. The results show that: when we used the whole denoised spectrum to construct model, the prediction effect of DBN is better than that of PLS, R2 is 0.879 4, RMSEP is 0.060 3 and RSD is 2.18%. When we used the selected characteristic band to model, the prediction effect of the PLS model is better than that of the DBN model. Using the 27 optimized characteristic wavelength variables to construct model, DBN has a good prediction effect, R2 is 0.958 4, RMSEP is 0.035 0 and RSD is 1.31%. It shows that the generalization ability of DBN is better, which achieved better prediction results by using a small number of wavelength variables. The proposed method in this paper can meet the practical needs, and provide technical support for online detecting and regulating TFAs content and producing low/zero TFAs oil products.