论文部分内容阅读
电子鼻漂移是气敏传感器的固有行为,用空载数据揭示漂移现象更具有一般性。为了有效去除电子鼻漂移,提出了一种基于空载条件下与小波包分解的漂移去除方法。对电子鼻空载数据进行小波包分解,获得小波包分解的逼近系数集;在对其进行离散度分析之后,构建了空载条件下的一种阈值函数。在此阈值函数基础上,扩展成为样本(有载)条件下的去漂移阈值函数,进而发展成有载样本的漂移剔除方法。为了检验该方法的有效性及实用性,将其应用于4种白酒的鉴别中。对4种白酒电子鼻数据按测试时间顺序生成训练集和测试集,线性的Fisher判别分析(FDA)结果表明,训练集、测试集数据处理前后的鉴别正确率均得到了提高,最低提高值为23.65%。这表明此方法能够提升电子鼻的检测能力。同时,为了进一步检验该漂移去除方法的性能,采用非线性的BP神经网络进行鉴别分析,结果显示:训练集的鉴别正确率从处理前的65.5%提高到处理后的100.0%,处理后的测试集鉴别正确率也达到了97.5%。这不仅说明了4种白酒的鉴别属较复杂的非线性分类问题,还充分说明了该漂移去除方法的有效性。
Electronic nose drift is the inherent behavior of gas sensors. It is more general to reveal drift phenomena with no-load data. In order to effectively remove the electronic nose drift, a drift removal method based on wavelet packet decomposition under no-load condition is proposed. The electronic nose no-load data is decomposed by wavelet packet to obtain the approximation coefficient set of wavelet packet decomposition. After analyzing the dispersion, a threshold function under no-load condition is constructed. On the basis of this threshold function, it expands to a function of de-drift threshold under the condition of sample (load), and then develops a drift-removing method of the loaded sample. In order to test the validity and practicability of the method, it was applied to the identification of four kinds of liquor. The training set and the test set were generated according to the testing time order of four kinds of liquor electronic nose data. Linear Fisher Discriminant Analysis (FDA) results show that the accuracy rate of the training set and the test set data before and after the data processing has been improved, the minimum increase is 23.65%. This shows that this method can enhance the detection ability of electronic nose. At the same time, in order to further test the performance of the drift removal method, the non-linear BP neural network is used for identification analysis. The results show that the discrimination accuracy of the training set is improved from 65.5% before treatment to 100.0% after the treatment, Set correct identification rate reached 97.5%. This not only shows that the identification of four kinds of white wine is a more complicated nonlinear classification problem, but also fully demonstrates the effectiveness of the drift removal method.