Isolating the Contributions of Resistance-Ratio Normalization and Humidity Fusion in a Minimal Three-Sensor...

Low-cost metal oxide semiconductor (MOS) electronic noses often combine resistance-ratio (R_s/R₀) normalization with a humidity transducer, both from the outset, so neither is measured. We isolate each on a three-sensor MQ array across nine pure headspaces. On one device the voltage-to-resistance step is monotone, so on 245 trials with per-trial clean-air baselines, the whole effect is the R₀ division: it cost a random forest 3.8 percentage points (seed-sweep interval 2.3-5.3) and linear models 8-9 points, possibly inflated by incomplete baseline recovery, and left the dominant error mode, a symmetric glycerol-water confusion, untouched. In-chamber humidity, from a channel withheld from the classifier, points to a physical rather than signal-conditioning origin. Admitting humidity as a fourth feature raises accuracy from 90.2 to 99.3% (450 trials, ten seeds, p < 10^-8) and from 86.7 to 93.3% under a cross-day holdout; temperature and equivalent CO₂ (eCO₂) do not. A leave-one-sensor-out test shows MQ-9 adds 0.89 ± 0.63 points on three channels and nothing with humidity admitted, so two MOS channels and humidity suffice here. The contribution is the matched ablation, not the accuracy: we propose, for MOS arrays sampling aqueous preparations, adding a transducer for the interferent rather than transforming channels that cannot resolve it.