Disentangling Response Magnitude and Relative MOS Sensor Fingerprints for Electronic-Nose Screening of Off-Odor in...

A multichannel metal-oxide semiconductor (MOS) electronic-nose response contains complementary information on overall response magnitude and the relative pattern among sensors. The most appropriate representation therefore depends on whether the aim is to distinguish volatile profiles, response levels, or both. This study compared three mathematically linked representations of the same six-channel NASYS S3+ response: global magnitude (M), centered log-ratio profiles (CLR), and per-sensor log-conductance ratios (LGR), which retain both components. Reference solutions varied both compound identity and concentration (four compounds, three levels, six sessions). The same framework was applied to 24 commercial cans of Californian-style black table olives preassigned by the supplier as conforming (C), mild (DS), or strong (DF) barnyard-like off-odor: 15 for calibration, six reserved before fitting, and three coded unknowns. Using the same linear discriminant analysis (LDA), CLR retained reference-compound class information better than M (balanced accuracy, BA, 0.83-0.89 vs. ≤0.56). In olives, grouped cross-validation favored M (BA 0.900), whereas LGR classified all six reserve cans and M only four; both results are descriptive within-campaign comparisons on 15 and 6 cans, not external validation. The three coded unknowns matched the concealed supplier labels. Signal representation should therefore be selected according to the analytical question rather than solely by calibration cross-validation performance.