电子鼻在长期使用过程中，由于受到传感器老化和温湿度波动等因素的影响，使得电子鼻信号存在着漂移现象，制约了电子鼻的准确检测。为了提高电子鼻鉴别多组分复杂食品时的长期稳健性，提出了一种独立分量分析（independent component analysis，ICA）融合小波能量阈值的去漂移信号方法。首先，运用ICA分解电子鼻信号以获得各个独立成分；其次，根据各个独立成分的小波能量值选择独立成分，以剔除漂移信号；最后，运用所选择的独立成分重构电子鼻信号。选用6种白酒样品和6种食醋样品进行电子鼻鉴别验证：在分别提取2类样品去漂移前后的电子鼻信号积分值（integral value，INV）特征后，分别对6种白酒样品和6种食醋样品进行了Fisher判别分析（fisher discriminant analysis，FDA），FDA的鉴别正确率分别由去漂移前的34.3%（白酒样品）、75.7%（食醋样品）提升到去漂移后的100%（白酒样品）、99.7%（食醋样品）。Fisher判别分析结果表明，所提出的电子鼻信号去漂移方法是有效的。与其他独立分量分析方法相比，该方法不受先验信息的影响，更便于实际应用，研究结果不仅适用于白酒、食醋样品的鉴别，也为其他类型的样品鉴别提供参考。%In the long-term application of the electronic nose (e-nose), drift of gas sensor, which is caused by sensor aging, environmental variations such as temperature and humidity, and other factors, will remarkably decrease the correct identification rate of the e-nose. In order to enhance long-term robust discriminant ability of the e-nose applied to food discrimination, a drift elimination/compensation method based on independent component analysis (ICA) coupled with wavelet energy threshold was proposed in our investigation. For the proposed method, firstly, a fast fixed-point algorithm for ICA (FastICA) based on non-Gaussianity maximization of negentropy was used to decompose the e-nose signals and generated many corresponding independent components. Secondly, some independent components were selected by their wavelet energy with the help of wavelet analysis so as to eliminate those independent components corresponding to drift signals. The reason is that some independent components corresponding to the constituents of food samples should be mainstream, and they can reflect the quality characteristic of food, and their energy values are relatively larger than that of the independent components corresponding to drift signals, then these independent components can be selected by wavelet energy threshold value. Namely, the independent components for which wavelet energy values are greater than the threshold value will be selected, otherwise will be removed, and the removed independent components may be considered to correspond to the e-nose drift signals. Finally, the e-nose signals which did not contain drift signals were obtained by reconstructing these selected independent components so as to carry out subsequent discrimination works. In order to test the validity of the proposed method, six kinds of white spirit samples and six kinds of vinegar samples were selected as identification objects. These white spirit samples (or vinegar samples) belong to three product groups, and each product group has two grades samples which are close in quality so as to increase the degree of difficulty of discrimination work. By trial and error, the wavelet energy threshold value for white spirit samples and vinegar samples were 0.075 and 0.258, respectively. After the e-nose signals of white spirit and vinegar samples were handled by the proposed method and the corresponding reconstructed e-nose signals were also obtained, the integral values (INV) selected as a kind of feature of the original e-nose signals and the reconstructed e-nose signals could be extracted. When Fisher discriminant analysis (FDA) was employed to deal with these features data, the correct identification rates of white spirit and vinegar samples increased from 34.3%(white spirit) and 75.7% (vinegar) up to 100% and 99.7%, respectively, and the cross-validation rates also increased from 35.0% and 74.3% up to 100% and 98.6%, respectively. The FDA results clearly show the proposed method is very effective, and the long-term robust identification ability of the e-nose for white spirit and vinegar samples was significantly enhanced. In addition, compared with other ICA algorithm, the proposed method does not require prior information, so it is a very simple method and more suitable for practical application. Furthermore, we think the method also has the reference value to the long-term robust identification of other food samples by e-nose.
展开▼
