Time-Conditioned Baseline Drift Compensation for Electronic Nose-Based Air Quality Monitoring Using Extreme Learning Machine

Electronic noses and MOS gas sensor arrays are promising for low-cost air quality monitoring, but long-term deployment suffers from baseline drift caused by environmental variation and sensor aging, leading to severe cross-time distribution shifts and performance degradation. This paper proposes an efficient drift-compensation framework based on Extreme Learning Machine (ELM) and semi-supervised domain adaptation. We employ Domain Adaptation ELM (DAELM) to jointly use labeled source data and a small set of target samples with closed-form training, and further propose CDAELM by adding a covariance-fusion regularization term that penalizes second-order statistic mismatch via the Log-Euclidean distance on the SPD manifold. Experiments on the 36-month UCSD gas sensor drift dataset under long-term deployment and short- term incremental-update protocols. Relative to conventional ELM, the enhanced ELM improves average recognition accuracy by 1.6%, while DAELM and CDAELM yield larger gains of 9.12% and 10.50%, respectively. Overall, these results demonstrate improved accuracy and robustness, supporting reliable long- term air quality and volatile organic compound (VOC) monitoring.