结合太赫兹时域光谱技术和支持向量机对3种典型产地的咖啡豆进行了鉴别.选取埃塞俄比亚(Ethiopia)、哥斯达黎加(Costa Rica)以及印度尼西亚(Indonesia)3个产地咖啡豆样品进行压片处理,采用太赫兹透射模式获取样品的时域和频域光谱信号,并用主成分分析法对太赫兹频域光谱信号进行分析;构造了基于粒子群(partical swarm optimization,PSO)参数寻优的支持向量机(support vector machine,SVM)鉴别模型,模型对不同产地咖啡豆样品的综合识别正确率达到95%.试验结果表明,太赫兹作为新型的检测手段结合模式识别方法可用于咖啡豆的产地鉴别.该文为一类在太赫兹波段下没有明显特征吸收峰的农产品/食品安全检测和产地追溯研究提供了一种快速、准确的方法.%Coffee is a very popular beverage in many countries. Coffee bean from different producing area has different flavour and functional properties, and thus the identification of producing area of coffee bean is important to assure the quality of coffee bean. The feasibility of a rapid and precise determination method of producing area of coffee bean was examined by using the terahertz (THz) time-domain spectra system (TAS7500TS HF1, Advantest Co., Ltd, Japan). Coffee bean samples from 3 different typical producing areas (Ethiopia, Costa Rica, and Indonesia) were collected and pressed into pellets for THz measurements. A total of 180 pellet samples (3 classes, each had 60 pellet samples) were randomly divided into calibration set (40 pellet samples for each class) and prediction set (20 pellet samples for each class). THz time-domain spectroscopy system worked with the TAS7500TS equipment in transmission mode. Before the experiment, the dry air was injected until the relative humidity reached below 3% to reduce the absorption of the THz waves by water in air. The parameters of THz system were as follow: frequency range was from 0.1 to 4 THz, the resolution was 7.6 GHz, the short pulse width was less than 50 fs and the average power was 20 mW. For each sample, the THz time-domain spectra were measured for 3 times at different position and then the average values were obtained. The frequency-domain spectra were acquired by a fast Fourier transform (FFT). Principal component analysis (PCA) with frequency-domain spectral data was performed to examine the qualitative difference of these 3 classes of coffee beans using the first 3 score vectors. The 3 groups of different class of coffee beans were almost apart from each other in the space of the first 3 principal components (PCs), although there was some overlap among the groups, which may be due to that the first 3 PCs only accounted for the all spectral variations of 68.75%. Thus, to reduce the dimension of the model features and retain more information of the THz spectra of samples, the first 20 components were selected as the spectral characteristics for the determination of producing area of coffee bean. The support vector machine (SVM), as a learning algorithm used for classification and regression tasks, was used to get the identification model. During the iteration for the optimum parameters selection, the particle swarm optimization (PSO) was designed, which could enlarge search space and improve search efficiency. The identification results of the PSO-SVM were compared with the least squares -support vector machine (LS-SVM) and back propagation neural network (BPNN). From the comparison, it was showed that the discrimination accuracy of all 3 classes of coffee beans using the PSO-SVM was up to 95% in prediction set and 100% in calibration set, respectively, which was the best model among the 3 methods. It can be concluded that the THz frequency spectra can be used as important features to identify the producing area of the coffee bean. The model with SVM method based on PSO can get better parameters of SVM to improve the identification ability than the traditional LS-SVM. THz spectra system combined with the proposed algorithm has been proved to be a very powerful and attractive tool for identification of producing area of coffee bean.
展开▼
