A drift correction method of E-nose data based on wavelet packet decomposition and no-load data: Case study on the robust identification of Chinese spirits

Abstract

Due to sensor aging, environmental changes and other factors, the drift of electronic nose (E-nose) signal is inevitable, and make E-nose does not own long-term robust detection capability. In order to improve the long-term detection capability of E-nose, a drift correction method based on wavelet packet decomposition and no-load data acquisition is proposed. Firstly, a no-load threshold function (NLTF) was proposed by decomposing the no-load data of the E-nose using wavelet packet decomposition, and then the NLTF was converted into a threshold function suiting for sample data (i.e. sample threshold function, STF). Secondly, Based on a concept of “sample measurement time window” (SMTW), the STF was employed to process the sample data within the SMTW; so that the drift contained in the sample data could be corrected. Finally, when the SMTW was recursively moved forward, the drift in all sample data corresponding to different time (or SMTW) could be corrected. As a study case, to realize the long-term robust detection of 6 kinds of Chinese spirits, the six kinds of Chinese spirits samples were tested intermittently for 12 months. When the SMTW was 3 months and the SMTW moved recursively forward 1 month every time, and after the above-mentioned drift correction method was applied to deal with these samples data within the SMTW, a long-term robust detection model based on Fisher discriminant analysis (FDA) was constructed with help of the idea of recursive correction. The model was able to carry out long-term robust detection for the six spirits samples; the correct identification rate could reach 100%. In addition, we also believe that the drift correction method has certain reference value for other E-nose data.