Vergara Gas-Sensor Dataset Overview

• The Vergara Gas-Sensor Dataset is a comprehensive benchmark featuring time-resolved resistance measurements from MOx sensor arrays in controlled wind-tunnel experiments.
• It offers multiple experimental configurations—including full-panel and drift-minimized subsets—to support robust gas classification, drift analysis, and domain adaptation studies.
• Advanced methodologies like DRCA, Knowledge Distillation, and SNM-Net are applied to address sensor baseline drift, ensuring realistic evaluation and improved cross-batch performance.

The Vergara Gas-Sensor Dataset is a widely used, publicly available benchmark for evaluating gas classification algorithms in the domains of artificial olfaction and electronic-nose (E-nose) research. Comprising extensive, time-resolved resistance measurements from large metal-oxide (MOx) sensor arrays exposed to diverse target analytes under controlled wind-tunnel conditions, the dataset is central to studies on drift, robust recognition, and domain adaptation in chemosensory systems. Its detailed experimental protocols, physical structure, and statistical properties have been both foundational and the subject of methodological scrutiny in recent literature (Dennler et al., 2021, Lin et al., 22 Jul 2025, Chen et al., 28 Dec 2025).

1. Experimental Design and Data Structure

The dataset was recorded over 16 months using an array of nine identical modules, each equipped with eight Figaro MOx sensors—after excluding consistently noisy channels, seven sensor types per module (TGS 2610, 2611, 2612, 2600 × 2, 2602, 2620 × 2) remain (Dennler et al., 2021). Sensors report resistance (read at 12-bit resolution) transformed via $$R_{sensor} = 10\ \mathrm{k}\Omega \cdot (3.11\,V - V_{sensor}) / V_{sensor}$$, reflecting redox reactions on heated metal-oxide surfaces.

A wind tunnel ($$2.5\,\mathrm{m}\,\times\,1.2\,\mathrm{m}\,\times\,0.4\,\mathrm{m}$$) delivers three wind speeds (0.10, 0.21, 0.34 m/s), with sensor boards placed at six discrete downstream points (P1–P6). Each trial involves one of ten analyte gases—acetone, acetaldehyde, ammonia, butanol, ethylene, methane, methanol, carbon monoxide, benzene, and toluene—delivered at fixed concentrations (typically 100–1000 ppm, up to 10 000 ppm for ammonia) for a 180 s exposure within a 260 s trial (gas injection: $t=20$ s to $t=200$ s). Five heater voltages (4.0 V–6.0 V) modulate sensor temperature. Each gas/position/wind/voltage variant is repeated 20 times, producing $\sim$18,000 trials, each sampled at 100 Hz.

Data for each trial are stored with filenames embedding precise timestamps. A key observation is that measurements for a given gas (and experimental group) are temporally clustered ("batched") rather than randomized, resulting in strong correlations of pre-exposure sensor baselines with gas identity (Dennler et al., 2021, Chen et al., 28 Dec 2025).

2. Dataset Variants and Derived Benchmarks

Operationally, two major data layouts are prevalent:

Dataset Variant	Sensor Configuration	Gases (Classes)	Sample Structure	Use Case
Full Vergara Dataset	9 boards × 8 sensors	10	Time series (260 × 72) per trial	Drift, closed-set recognition (Chen et al., 28 Dec 2025)
UCI Drift Dataset	16 MOx sensors (array)	6 (from above)	Matrix: $\mathbb{R}^{N\times 16}$	Drift, domain adaptation (Lin et al., 22 Jul 2025)
Drift‐Minimized Subset	Board 3, sensors 2,3,5–8	3 (methanol, ethylene, butanol)	~1200 trials, focused in time	Conservative benchmarking (Dennler et al., 2021)

The UCI Gas Sensor Array Drift Dataset (taken from Vergara et al.) contains 16 sensors and six chemicals, arranged in ten chronological batches spanning 36 months (Lin et al., 22 Jul 2025). Each sample is a 16-dimensional resistance vector and a categorical label $\in\{1,\dots,6\}$. This variant is especially used for drift-compensation and domain adaptation analyses.

3. Drift Phenomena and Statistical Characterization

The dataset exposes both short-term and long-term drift mechanisms:

• Baseline drift: The mean resistance prior to gas exposure ($t\in [0,20\,\mathrm{s})$) varies between trials and escalates over months.
• Short-term intra-trial drift: Sensor responses can fluctuate within a single trial.
• Long-term inter-trial drift: Changes in pre-exposure baseline span across months or batch boundaries.

Quantitatively, drift is measured using the coefficient of variation $c_v = \sigma/\mu$ ("cv") across trials (long-term) and within-trial segments (short-term). Spatial cv maps reveal sensors near the plume center, and in particular channel 4 (TGS2602), exhibit elevated drift.

Instantaneous drift is defined per sensor $i$ and time $$2.5\,\mathrm{m}\,\times\,1.2\,\mathrm{m}\,\times\,0.4\,\mathrm{m}$$0 as $$2.5\,\mathrm{m}\,\times\,1.2\,\mathrm{m}\,\times\,0.4\,\mathrm{m}$$1. "Zero-offset subtraction" mitigates slow baseline changes:

$$2.5\,\mathrm{m}\,\times\,1.2\,\mathrm{m}\,\times\,0.4\,\mathrm{m}$$2

with $$2.5\,\mathrm{m}\,\times\,1.2\,\mathrm{m}\,\times\,0.4\,\mathrm{m}$$3 averaged over $$2.5\,\mathrm{m}\,\times\,1.2\,\mathrm{m}\,\times\,0.4\,\mathrm{m}$$4. However, residual drift persists, especially before gas arrival, and remains informative about batch identity.

Strong temporal correlations in drift are evidenced by abrupt baseline discontinuities aligning with gas batch transitions (Dennler et al., 2021). This batch effect constitutes a confounding variable in supervised learning: models may infer gas type by exploiting baseline drift rather than true analyte-induced signal features.

4. Approaches for Drift Mitigation and Recognition

Recent methodologies for addressing sensor drift and robust gas recognition leverage both statistical and deep learning frameworks:

• Domain Regularized Component Analysis (DRCA): Projects features into a subspace optimizing within-domain similarity and between-domain separation for source and target batches. The objective balances intrabatch scatter and penalizes batch-to-batch domain shift (Lin et al., 22 Jul 2025).
• Knowledge Distillation (KD): Employs a teacher–student neural network paradigm, where a teacher model trained on source data generates soft-labels (temperature-smoothed) for unlabeled target data. The student network jointly minimizes source cross-entropy and soft-label (KL-divergence) distillation loss, parameterized by:

$$2.5\,\mathrm{m}\,\times\,1.2\,\mathrm{m}\,\times\,0.4\,\mathrm{m}$$5

where $$2.5\,\mathrm{m}\,\times\,1.2\,\mathrm{m}\,\times\,0.4\,\mathrm{m}$$6 weights the distillation component.

• SNM-Net (Spherical Normalization + Mahalanobis Scoring): Implements geometric decoupling by cascading batch normalization and $$2.5\,\mathrm{m}\,\times\,1.2\,\mathrm{m}\,\times\,0.4\,\mathrm{m}$$7-normalization, projecting backbone features onto a unit hypersphere to remove magnitude drift. Class decision boundaries are formed using Mahalanobis distance with per-class mean $$2.5\,\mathrm{m}\,\times\,1.2\,\mathrm{m}\,\times\,0.4\,\mathrm{m}$$8 and (regularized) covariance $$2.5\,\mathrm{m}\,\times\,1.2\,\mathrm{m}\,\times\,0.4\,\mathrm{m}$$9:

$t=20$0

Unknown gas detection is performed via a threshold on minimal Mahalanobis distances, set as the 95th percentile on validation scores (FPR $t=20$1 5%) (Chen et al., 28 Dec 2025).

5. Benchmarking Protocols and Statistical Evaluation

Classification performance is systematically quantified using accuracy, precision, recall, and F1-score:

• $t=20$2
• $t=20$3
• $t=20$4
• $t=20$5

Protocol specifics include:

• For the full dataset: Linear SVM, 80/20 train/test, repeated 10 times, 100 ms windows (Dennler et al., 2021).
• For SNM-Net: 10-fold cross-validation, holding out four of ten classes as "unknown" per fold, evaluating accuracy, AUROC, and TPR@FPR=5% over L1–L5, averaging over 50 runs (Chen et al., 28 Dec 2025).
• For UCI drift: Two main tasks—train on batch 1, test on remaining batches ("one-to-many"), and cumulative training up to batch $t=20$6, test on batch $t=20$7 ("cumulative-to-next"), using a four-layer fully-connected network with extensive randomized splits and paired $t=20$8-test validation (Lin et al., 22 Jul 2025).

SNM-Net, using a Transformer backbone, achieved mean AUROC 0.9977 ($t=20$90.0028), known-gas accuracy $t=200$0, and unknown gas TPR@FPR=5% of $t=200$1, with position stdev $t=200$2 (Chen et al., 28 Dec 2025).

6. Dataset Limitations and Recommendations

Analysis revealed that temporally batched gas presentations in the original protocol induce strong time–gas correlations in pre-exposure sensor baselines. Even after zero-offset subtraction, models can exploit residual drift, often yielding artificially high classification scores (Dennler et al., 2021). A "minimally drift-affected" subset—comprising overlapping batches of methanol, ethylene, and butanol, board 3, and excluding high-drift sensors—yields more realistic 3-way accuracy ($t=200$360%) after offset subtraction.

Key recommendations for future gas-sensor datasets:

1. Pseudo-randomize gas presentation and settings to break time–identity correlations.
2. Insert frequent reference or clean-air trials for ongoing drift monitoring and correction.
3. When full randomization is infeasible, record multiple independent batches and apply leave-batch-out cross-validation.
4. Archive and share full raw data, including timestamps, humidity, temperature, and perturbations.
5. Disclose all drift analyses and preprocessing strategies in benchmark reporting.

Implementing these practices ensures that gas classification outcomes reflect true analyte separation rather than unintentional experimental artifacts.

7. Impact on Algorithm Development and Domain Adaptation

The Vergara dataset has catalyzed research in drift adaptation, robust open-set recognition, and benchmarking for E-nose systems. Studies have demonstrated that classical models, when trained and tested on data spanning drifted batches, can vastly overestimate generalization (Dennler et al., 2021). Domain adaptation protocols, as instantiated in tasks from (Lin et al., 22 Jul 2025), explicitly treat cross-batch differences as $t=200$4 domain shifts, formalizing the drift problem as statistical learning under distributional change.

New neural frameworks such as SNM-Net establish reference procedures for achieving accuracy–stability trade-offs under drift and unknown-analyte interference, employing geometric and statistical normalization strategies that have near parameter-invariant performance across sensor positions and experimental folds (Chen et al., 28 Dec 2025). This suggests that domain adaptation and out-of-distribution detection in E-nose datasets increasingly rely on rigorously characterized benchmarks set by the Vergara protocol.