Mitigating Sensor Drift in E-Nose Using Different Algorithms
ARTIFICIAL INTELLIGENCE, SMART ENGINEERING AND SUSTAINABLE ENERGY SYSTEMS: EMERGING TECHNOLOGIES FOR INDUSTRY 5.0, Assoc. Prof. Dr. Erdal ÇILĞIN, Editör, BZT TURAN PUBLISHING HOUSE, Delaware, ss.17-30, 2026
- Yayın Türü: Kitapta Bölüm / Diğer
- Basım Tarihi: 2026
- Doi Numarası: 10.30546/19023.978-9952-610-75-8.2026.100.1132
- Yayınevi: BZT TURAN PUBLISHING HOUSE
- Basıldığı Şehir: Delaware
- Sayfa Sayıları: ss.17-30
- Editörler: Assoc. Prof. Dr. Erdal ÇILĞIN, Editör
- Karadeniz Teknik Üniversitesi Adresli: Evet
Özet
Electronic noses (e-noses) are widely used for food classification; however, sensor drift between training and test data can significantly reduce performance. In this study, we investigated drift compensation techniques on a three-class food dataset consisting of Cake, Pastry and Fish samples collected in 2021 and 2022. These samples were measured at the 60th minute while being cooked in an oven set to 170°C. The e-nose consisted of an array of 8 MOS sensors, which were used to measure and capture the volatile compounds emitted by the food samples, providing the input data for the classification models. Two classification methods, Support Vector Machine (SVM) and Partial Least Squares Discriminant Analysis (PLS-DA), were employed to evaluate the classification performance of the models. To mitigate drift effects, three commonly used correction techniques—Mean Shift, Z-Score normalization, and CORAL (Correlation Alignment)—were applied. Both classifiers were trained on 2021 data and tested on 2022 data. Baseline results without drift correction achieved test accuracy of 75% and 80% for SVM and PLS-DA, respectively. For SVM, test accuracy increased to 85% with both Mean Shift and Z-Score methods, while CORAL achieved an accuracy of 80%. Additionally, for PLS-DA, test accuracy increased to 85% with Z-Score normalization, whereas Mean Shift reduced the accuracy to 75%, and CORAL maintained an accuracy of 80%. These results demonstrate that simple statistical drift correction methods can effectively mitigate sensor drift in e-nose applications, enhancing classification reliability. This approach is particularly suitable for small dataset and for scenarios requiring real-time correction.