Drift Elimination Method of Electronic Nose Signals Based on Wavelet Analysis and Discrimination of White Spirit Samples

Abstract:

In order to enhance the longterm identification accuracy and robustness of enose, a drift elimination method of electronic nose (enose) signals based on wavelet analysis was proposed. Firstly, wavelet decomposition was used to decompose the enose data contained drift and generated decomposition coefficients. Secondly, a relative deviation threshold filtering function was constructed to threshold wavelet coefficients and then the corrected wavelet coefficients were obtained. Finally, the enose signals which had less or not drift signals were obtained by reconstructing the corrected coefficients. For six kinds of discriminated white spirit samples, five groups of training set samples and corresponding test set samples which were randomly generated were carried out the drift elimination processing and signal reconstruction by the proposed method. After the integral values (INV) selected as a feature of the original/reconstructed enose signals were extracted, Fisher discriminant analysis (FDA) and BP neural network were employed to deal with these feature arrays of the five groups of data patterns. The FDA results clearly showed that the highest correct identification rate of five groups of training set and test set samples was 45% before drift eliminating and up to all 100% after drift eliminating, respectively. Meanwhile, BP neural network results also showed that the highest correct identification rate of the five group samples was 31.7% before drift eliminating, and the correct identification rate was up to 98.3% after drift eliminating. The two kinds of identification results illustrated the proposed method was very effective and robust for white spirit samples identification. In addition, the drift elimination method also had the reference value for the identification of other food samples.