基于独立组分分析和BP神经网络的可见/近红外光谱蜂蜜品牌的鉴别

提出了一种基于独立组分分析的可见/近红外光谱透射技术快速鉴别蜂蜜品牌的新方法。用独立组分分析方法获取蜂蜜的可见/近红外光谱载荷图,将载荷图中相关性最大的波段,作为人工神经网络的输入建立蜂蜜品牌的鉴别模型。建立了一个三层的BP神经网络模型,各层传递函数采用S型(Sigmoid)函数,并设置网络输入层节点数为9,隐含层节点数为10,输出层节点数为3。每个品牌25个样本,3个品牌共75个样本,用来建立BP神经网络模型,剩余的3个品牌各5个样本用于预测,鉴别准确率达100%,模型的拟合残差为8·245365×10-5。说明基于独立组分分析的方法具有很好的鉴别效果,为蜂蜜的品牌鉴别提供了一种新方法。

Application of Visible/Near Infrared Spectroscopy to Discriminating Honey Brands Based on Independent Component Analysis and BP Neural Network

Visible/near infrared spectroscopy (Vis/NIRS) appears to be a rapid and convenient non-destructive technique that can measure the quality and compositional attributes of many substances. In the present study, a nondestructive method for the classification of honey brands was developed using Vis/NIRS. The honey brands studied in the research were Feng boshi, Tian ranfeng and Guan shengyuan. The sample set comprised 30 of each brand. Independent component analysis (ICA) was put forwarded to select several optimal wavelengths based on loading weights. Two types of preprocessing (Savitzky-Golay combined with multiplicative scatter correction) were used before the spectral data were analyzed with multivariate calibration methods of artificial neural network (ANN). The absorbance values log (1/T) (T= transmission), corresponding to the wavelengths of 408, 412, 409, 1 000, 468, 462, 408, 400, 997 and 998 nm were chosen as the input data of ANN. The ANN model with three layers was built, and the transfer function of sigmoid was used in each layer. After several trials, the best neural network architecture was obtained with 10 nodes in hidden layers. In the model, the node of input layer, hidden layer, output layer was set to be 9, 10, and 3 respectively, and the goal error was set to be 0. 000 1, the speed of learning was set to be 0.2, the time of training was set to be 1 500. Seventy five samples (25 with each brand) from three brands were selected randomly as calibration set, and the left 15 samples (5 with each brand) were as perdition set. The discrimination rate of 100% was achieved, and the fitting residual was 8. 245 365 x 10(-5). These indicated that the result of honey discrimination was very good based on ICA method, and offer a new approach to the fast discrimination of varieties of honey.