E-nose drift correction method based on no-load data and its application of robust identification for identifying vinegar

Abstract: Electronic nose (E-nose) signal drift is inevitable due to sensor aging and fluctuation of ambient temperature and humidity, which could compromise its ability of robust long-term detection. To ensure long-term robust service of the E-nose for detecting vinegar samples, a drift recursive correction method is proposed in this paper using the wavelet packet decomposition coefficients based on no-load data. The method does not require special correction processing and the sensor drift can be corrected based only on the no-load response data and the sample response data of the E-nose. In the model, the Symlet wavelet function was first used to decompose the no-load data of the E-nose using a no-load threshold function (NLTF) given in the paper. The NLTF was then converted to sample threshold function (STF) suiting the samples data by a constructed adjustment coefficient. Using the STF, a correction function based on the wavelet packet decomposition coefficient of the samples of E-nose data was constructed. The E-nose data of six vinegars were subject to a drift correction by means of the correction function. We also introduced the concept of "sample measurement time window" (SMTW), and used the correction function to process the sample data within the SMTW. As the SMTW progresses recursively, the drift in all sample data at different times (or SMTW) could be used to recursively correct the samples of the vinegars. To validate the drift correction method and test the applicability of the SMTW, the sample data in the SMTW were used as a training set and the sample data between one month and two months after the SMTW were used as test set. A recursive Fisher discriminant analysis (FDA) model was built, which was proven capable of long-term robust detection of the vinegars. The samples of the vinegars were tested intermittently for 16 months, and the SMTW in which was 6, 5, 4 and 3 months, respectively. With the change in SMTW, the correct discrimination rate for the training set and the test set also changes. When the SMTW was more than 4 months or less than 4 months, the correction identification rate of FDA was less than 100%, and the correction identification rate was only 92.22% under certain circumstance. Therefore, when the SMTW was 6, 5 or 3 months, the samples of the vinegars cannot be identified robustly in long term. When SMTW was 4 months, the test samples in SMTW and the samples within one month after the SMTW were effectively identified by the established recursive FDA model, and the vinegar samples can be identified robustly in long term, with a correction identification rate of 100%. That is, the test samples within one month after the SMTW could be accurately identified using the FDA model built from the sample E-nose data when the SMTW was within 4 months. We believe that our results has implications as the proposed method is applicable to other E-nose data.