Mo-SML-02 - Progressive Domain Adaptation under Sensor Drift for Gas Classification with Limited Calibration Labels

Event

EUROSENSORS 2026
2026-09-06 - 2026-09-09
Zurich

Band

Poster

Chapter

Signal Analysis, Machine Learning, Physics-Informed Machine Learning And Artificial Intelligence For Sensing

Author(s)

H. Fan, A. J. Lilienthal - Technische Universität München,Munich (DE) and Örebro University,Örebro (SE)

Pages

690 - 691

DOI

10.5162/eurosensors2026/Mo-SML-02

ISBN

978-3-910600-12-6

Price

free

Abstract

Metal-oxide electronic noses suffer from sensor drift, in which sensor response characteristics change over time. This creates a data distribution shift (domain gap) that severely degrades gas classification accuracy. To address this drift effect, domain adaptation techniques reduce the distribution discrepancy between the initial source domain and drift-affected target measurements. We study batch-wise adaptation across temporally ordered drift batches, where calibration labels are extremely scarce. To ensure robust adaptation under unpredictable drift patterns, we train a diverse pool of conditional domain-adversarial experts. We then fuse their predictions using a confidence-aware Expectation-Maximization procedure. This fusion mechanism acts as a smart consensus, estimating each expert's reliability using the limited calibration labels without requiring additional annotations. On the UCI drift benchmark, this budget-respecting approach consistently improves target accuracy over a tuned Drift Correction Autoencoder (DCAE) baseline, particularly on later, more severely drifted batches.