Vergara Gas-Sensor Benchmark

• Vergara Dataset is a collection of open datasets, notably a benchmark gas-sensor corpus widely used in artificial olfaction and open-set gas recognition communities.
• It employs a metal-oxide sensor array in a controlled wind tunnel setup to capture high-density, transient responses from 10 distinct gases under varied experimental conditions.
• Extensive analyses on sensor drift, preprocessing protocols, and open-set recognition highlight both the dataset's value and limitations for robust, drift-compensated gas classification.

The term "Vergara Dataset" refers to three prominent, technically distinct open datasets in contemporary research: (1) the gas-sensor array benchmark for electronic nose systems, (2) the “SANRlite” Spanish notary records for historical language modeling, and (3) the VerSe computed tomography vertebral segmentation dataset. Of these, the best established and most widely cited is the gas-sensor dataset by Vergara et al., which is the canonical reference in the artificial olfaction and open-set gas recognition communities. This article provides a comprehensive technical overview of this gas-sensor dataset, its design, known artifacts, and its foundational role as a benchmark in gas recognition research, along with briefer clarifications of its distinctness from the SANRlite (historical Spanish) and VerSe (vertebrae segmentation) datasets.

1. Dataset Acquisition and Physical Configuration

The Vergara gas-sensor dataset, first published by Vergara et al. and subsequently scrutinized in benchmark and drift analysis studies (Dennler et al., 2021), consists of detailed recordings from a metal-oxide (MOx) sensor array exposed to highly controlled pulses of various analyte gases. Data were collected in a purpose-built wind tunnel (2.5 m × 1.2 m × 0.4 m) under computer-supervised gas injection and purging. The sensor array comprises 72 MOx devices—distributed as nine identical modules (each eight sensors: models TGS 2611, 2612, 2610, 2602, 2600 ×2, and 2620 ×2)—arranged at six downwind locations (P1–P6). The setup enabled systematic variation of sensor-gas distance, airflow velocity (0.10, 0.21, and 0.34 m/s), and sensor-hotplate voltage (4.0–6.0 V). This meticulous experimental control yields high-density, repeatable transient response data under both nominal and drift-affected conditions.

2. Analytes, Protocol, and Temporal Structure

The dataset targets 10 "high-priority" gases: acetone, acetaldehyde, ammonia, butanol, ethylene, methane, methanol, carbon monoxide (CO), benzene, and toluene. Each is tested at characteristic concentrations (e.g., acetone, acetaldehyde, ammonia at 100 ppm; butanol at 50 ppm; ethylene at 200 ppm; benzene and toluene at 20 ppm). For each unique parameter combination (gas, concentration, board location, wind speed, hotplate voltage), 20 repeated trials are acquired. Each trial spans 260 s: the gas is injected from t=20 s to t=200 s; prior baseline and recovery periods bracket this window.

Data are acquired at a native 100 Hz sampling rate with 12-bit resolution, resulting in 26,000 samples per sensor per trial. Sensor outputs (voltages) are converted to resistance via the transformation $$R_{\text{sensor}} = 10\,\text{k}\Omega \times \frac{3.11\,\text{V}-V_\text{sensor}}{V_\text{sensor}}$$ enabling standardized downstream feature processing.

Critically, batch structure is non-randomized: all 300–400 trials of a given gas/concentration are run consecutively before proceeding to the next, a property with substantial implications for drift and classification artifact analysis (Dennler et al., 2021).

3. Drift Characterization and Benchmarking

The Vergara dataset is distinguished by the presence of both long-term and short-term sensor drift, extensively analyzed in later works (Dennler et al., 2021). Long-term drift is evidenced by session-wise step changes in pre-exposure sensor baselines, tightly coupled to batch (gas) identity. Short-term drift manifests within a single 260 s trial, even before gas exposure. Quantitatively, baseline coefficients of variation ($c_v$) can exceed 10–20% in some sensor/location combinations.

Compensation for drift typically involves zero-offset subtraction:

$\hat x_i(t) = x_i(t) - x_i(t_0)$

with $t_0=0$ or as the mean over the first 100 ms, aligning all trials to a common baseline (Dennler et al., 2021, Chen et al., 28 Dec 2025). Residual drift remains after this correction, leading to inflated classification results if not recognized.

Benchmark analyses using linear SVMs show that, without offset correction, the sensor baseline alone enables 10-way gas classification at $\sim$94% accuracy even before gas arrival ($t\in[0,0.1\,\rm s)$). After zero-offset subtraction, classification accuracy at $t=0$ subsides to chance ($10\%$ for ten classes), but residual drift supports non-chance classification until full gas response develops, with up to $\sim$80% accuracy by $t=35$ s. Only in a carefully selected, minimally drifted subset (methanol, ethylene, butanol; "board 3"; certain sensors excluded), does true gas-response-driven classification emerge, yielding $\sim$60% accuracy post-onset—a dramatic reduction relative to the unfiltered dataset.

4. Preprocessing and Experimental Protocols

Recent benchmark studies employ standardized preprocessing to ensure comparability and mitigate drift artifacts (Chen et al., 28 Dec 2025):

• Temporal downsampling: 100 Hz $\rightarrow$ 1 Hz (block average), yielding 260 time steps.
• Channel-wise normalization: $z_{ij} = (x_{ij} - \mu_j)/(\sigma_j + \varepsilon)$ for per-channel mean $\mu_j$ and standard deviation $\sigma_j$.
• Reshaping: Outputs cast to $260 \times 72$ (time $\times$ sensor) spatiotemporal arrays for CNN, RNN, or Transformer-based modeling.

Open-set recognition protocols are a core evaluation paradigm: for each of five locations (L1–L5), 10 random folds are evaluated, with six gases designated "known" (used for training, 60% of their samples), the remaining four "unknown" (used only in testing, 40% of known plus all unknown). Metrics are averaged over 50 cross-validation trials (Chen et al., 28 Dec 2025).

5. Open-Set Recognition, Drift Mitigation, and State-of-the-Art Results

The SNM-Net framework (Chen et al., 28 Dec 2025) establishes the current methodological baseline for open-set gas recognition on this corpus. Its feature normalization stack (batch normalization plus L2 spherical normalization) projects network outputs onto the unit hypersphere $S^{d-1}$, explicitly decoupling direction (chemically relevant) from intensity (drift-prone). Class centers and Mahalanobis distances on this hypersphere are used as adaptive rejection scores for unknown detection.

Empirical results show that the Transformer+SNM achieves near-theoretical open-set discrimination, with AUROC = 0.9977 and unknown detection rate (TPR at 5% FPR) of 99.57%. This represents a 3.0% improvement in AUROC and a 91% reduction in standard deviation compared to the prior best (Class Anchor Clustering), with robustness to sensor position evidenced by standard deviations $<0.0028$ across positions. Signal-intensity drift across positions induces order-of-magnitude feature norm variability—a challenge directly addressed by the normalization pipeline.

6. Dataset Limitations and Impact on Benchmarking

Perhaps the most critical finding regarding the Vergara dataset is that its temporally clustered batch structure conflates gas identity and measurement session, permitting trivial exploitation of long-term drift for "classification" (Dennler et al., 2021). This has led to widespread overestimation of model accuracy in previous research: nearly perfect closed-set classification accuracies can be achieved by leveraging baseline differences, rather than true chemical discrimination. Even after offset correction, short-term drift artifacts persist. Only a minority subset, carefully selected for minimal batch effects and drift, supports meaningful, gas-response-dominated classification, and then at substantially reduced accuracy compared to the full dataset.

A plausible implication is that the Vergara dataset, while exceptionally rich and valuable for drift analysis, diversity, and open-set protocol design, should be used with extreme caution as a closed-set gas identification benchmark without explicit artifact controls. It remains, however, the de facto open benchmark for the development and evaluation of on-line, robust, drift-compensated electronic nose algorithms (Chen et al., 28 Dec 2025).

7. Distinction from Other “Vergara” Datasets

The term "Vergara Dataset" is sometimes used informally for unrelated datasets:

• SANRlite/"Spanish Notary" Dataset: Historical text corpus from notarial records of Estenban Agreda de Vergara, comprising 162 pages of 17th-century Spanish, with high-resolution images, manual transcription (952 sentence-segments), and rich metadata for NLP fine-tuning. Empirically improves fine-tuned Spanish LLM performance over pre-trained or ChatGPT baselines, especially for classification and MLM tasks (Sarker et al., 2024). This dataset is intellectually and technically unrelated to the gas-sensor corpus.
• VerSe Vertebral Segmentation Dataset: Sometimes cited in similar transliteration as "Vergara," this resource is a large-scale CT dataset for anatomical vertebral segmentation, distinct in scope, modalities, and research community (Liebl et al., 2021).

Correct identification of the intended "Vergara Dataset" is essential to ensure methodological rigor and appropriate citation across signal processing, NLP, and medical imaging domains.