LAUSR

Low-concentration samples play an important role in the training of the electronic nose (E-nose), while low-concentration samples are difficult to obtain because of the limited accuracy of sensor detection, and the confect precision of low-concentration samples are hard to control. Traditionally, we can improve the accuracy of the sensors by improving their materials, but this method is costly and difficult to operate. We devised another efficient and convenient method by using algorithms, which can improve the recognition accuracy of low-concentration samples. In this paper, we put forward a novel technique that combines the global and local features, that is, the fusion of the features can reflect the characteristics of the data from comprehensive perspectives. We used this technique to explore whether there exists a situation that the recognition rate of E-nose trained by the samples with different concentration are the same or not. After that, we proposed a new evaluation index based on the dependence of low-concentration samples to judge kinds of feature extraction algorithms. The experimental results show that whether the algorithms of global features (PCA and ICA), the algorithms of local features (LPP and NPE) or the fusion, there are effective and alternative solutions, that is, we can use more high-concentration samples to replace the comparatively minor number of low-concentration samples in the training set. In addition, compared with other algorithms, the fusion of ICA and LPP based on weighted enumeration and EQPSO has the highest recognition rate with the least number of high-concentration samples.