Implementation and Analysis of a Sensor-Driven Air Quality Monitoring System in Kitchen Environments: A Calibration-Aware Pilot Study

2.1 Low-cost sensor calibration and cross-sensitivity

Low-cost gas and particulate sensors (the MQ-series, CCS811, GP2Y1010, Bosch BME680, etc.) are widely used in IoT-based IAQ monitoring [6-9], yet they are known to require careful calibration and cross-sensitivity handling. Ródenas García et al. [10] catalogue calibration, reliability and standardisation as the principal open issues for low-cost IAQ sensors. Sá et al. [6] arrive at the same conclusion in their review of low-cost sensing technology, noting that ppm-level claims from uncalibrated sensors should be treated as relative indices. Bousiotis et al. [5] co-located four low-cost particulate sensors with research-grade instruments and found that, while the relative trends were preserved, the absolute concentrations diverged substantially. These results frame the interpretation we adopt below: relative kitchen vs. indoor comparisons are defensible; absolute regulatory-grade concentrations are not.

Beyond cataloguing these limitations, the measurement-science literature prescribes how low-cost sensors should be field-calibrated and validated. Spinelle et al. [11] show that field calibration against co-located reference analysers—rather than laboratory or data-sheet curves alone—is a prerequisite for quantitative use of low-cost gas sensors, and Maag et al. [12] survey calibration strategies across air-pollution monitoring deployments, treating offset/gain correction, sensor fusion and periodic recalibration as the principal remedies for sensor drift. Drift and calibration transfer are first-order effects: Tancev and Pascale [13] show that field-calibrated sensor systems incur a systematic sampling bias when relocated from their calibration environment, and Karagulian et al. [14] consolidate field-evaluation results across commercial low-cost sensors, reporting device-to-device performance variation large enough to make validation deployment-specific. Within instrumentation practice, De Capua et al. [15] likewise propose strategies for improving the measurement quality of IAQ monitoring systems. This literature motivates the stance adopted here: with no co-located reference instrument available, the data-sheet calibration framework is documented in full and propagated to an explicit uncertainty statement (Section 3.3), but absolute concentrations are not claimed.

2.2 Practical meaning and limitations of eCO₂ and total volatile organic compound estimates

The CCS811 and similar metal-oxide gas sensors report on-chip algorithm-derived eCO₂ and total volatile organic compound (TVOC) values from a single sensor element. These outputs are useful trend indicators but do not measure CO₂ or speciated VOCs directly. Zhu et al. [16] explicitly treat CCS811-derived eCO₂ as an indicator for forecasting purposes rather than as ground-truth CO₂; the chip baseline is known to drift in sustained high-VOC environments without clean-air exposure. Shin et al. [17] note the same caveat in their monitoring work coupled to heating, ventilation and air conditioning (HVAC) systems. The TVOC value, in particular, is a composite, un-speciated proxy whose absolute magnitude is sensitive to humidity and to the chip’s internal compensation state.

2.3 Ventilation evaluation in kitchen environments

Kitchens are an under-studied but high-exposure subset of the IAQ monitoring problem space. Cooking events release CO, VOCs, NOₓ, and particulate matter at concentrations that can exceed outdoor levels; the effectiveness of exhaust hoods and ventilation openings depends strongly on flow rate, capture geometry, and occupant distance. Existing IoT-based IAQ deployments [7, 9, 18-21] either report a single-environment study without ventilation context or evaluate office and residential settings rather than institutional kitchens [22, 23]. The present pilot adopts a kitchen-versus-room comparison as a case study for the documented calibration framework above, rather than as the principal contribution. The interpretation is bounded by the limitations stated explicitly in Section 7: no air-change-rate measurement, no exhaust on/off comparison, and a single 24 h deployment per environment.

4.1 Environments and deployment

The system was deployed in two distinct environments to compare their air quality. The indoor (control) site was a standard student residential room without in-room cooking, without air conditioning, and with passive cross-ventilation through a single window. The kitchen site was a hostel-mess kitchen equipped with overhead exhaust hoods and wall-mounted air blowers, where approximately 200 meals per day are cooked on multi-burner liquefied petroleum gas (LPG) stoves; the adequacy of this ventilation under sustained cooking is the question we examine. The monitoring node was placed approximately 4.6 m (15 ft) horizontally from the nearest active burner, at a height of 1.5 m, on a stainless-steel preparation counter outside the immediate hood-capture zone.

4.2 Sampling, preprocessing and missing-value handling

Both deployments lasted approximately 24 h. The indoor deployment ran from 6 April 2024 12:00 IST to 7 April 2024 11:59 IST (1,333 samples; duration 23.99 h); the kitchen deployment ran from 5 May 2024 09:02 IST to 6 May 2024 08:50 IST (1,253 samples; duration 23.80 h). All timestamps are in Indian Standard Time (UTC + 05:30) as configured on the ThingSpeak channel. The median sampling interval was 62 s in the indoor deployment (IQR [61, 62] s) and 45 s in the kitchen deployment (IQR [42, 82] s); the wider kitchen IQR reflects occasional ESP8266 Wi-Fi reconnections. The longest single inter-sample gap was 37.0 min in the indoor deployment and 22.0 min in the kitchen deployment, in both cases attributable to transient Wi-Fi outages. The field-level CSV files contain no NaN values after the platform recorded the channel write; the analysis pipeline therefore treats inter-sample gaps as missing-by-design and performs no imputation. Hourly summary plots use a one-hour resample-by-mean window anchored at the top of each clock hour, computed via pandas DataFrame.resample('1h').mean(); all per-channel summary statistics in Section 6 operate on the raw per-sample data.

4.3 Data dictionary

Table 3 provides the data dictionary for the raw CSV files released with this study.

Table 3. Data dictionary for the released CSV files

6.1 Per-channel summary statistics and percentiles

Table 4 reports the per-channel summary statistics for the two environments. As anticipated under the calibration disclosure of Section 3.3, absolute values for MQ7, MQ135, and the dust channel are reported as ppm-equivalent or sensor-index, not as calibrated absolute concentrations. TVOC is consistently reported in ppb and eCO₂ in ppm throughout. The kitchen distributions are visibly more heavy-tailed for the chemistry channels (TVOC, eCO₂) than the indoor distributions, as expected for episodic cooking events. The indoor dust channel is the exception: its mean (14.8) and SD (72.2) are dominated by intermittent spikes rather than by a sustained level. Sixty of the 1,333 indoor samples (4.5%) exceed 50 and contribute about 83% of the summed signal, and 29 of the 30 samples above the kitchen maximum of 186 fall in the last 70 min of the deployment (10:52–11:57 IST on 07 April 2024), where the readings alternate between zero and values of up to 625 with no accompanying change in the other channels. The released data do not identify the source of this late cluster, so the readings are retained as recorded. Excluding the samples above 50, the indoor dust mean is 2.6 (SD 5.7). These spikes therefore raise the indoor mean, which implies that the kitchen-minus-indoor dust difference reported in Table 5 is, if anything, understated; consequently, the indoor median (0.0) is a more representative summary for this channel.

To ground the comparison quantitatively, Table 5 reports the kitchen-versus-indoor mean difference Δ for each channel alongside (i) the combined standard uncertainty u(Δ) propagated from the Table 2 accuracy bounds (Section 3.3) and (ii) the descriptive standardized effect size (Cohen’s d on the pooled per-sample SD). Six of the seven channels show |Δ| / u(Δ) > 1, with the chemistry and moisture channels well clear of the instrument bound—eCO₂ (≈8×), TVOC (≈15×), dust (≈8×) and humidity (≈4×)—so the observed kitchen–indoor contrasts on these channels cannot be attributed to manufacturer accuracy specifications alone. The MQ135 difference lies well within its propagated uncertainty (|Δ| / u(Δ) ≈ 0.1) and is therefore not interpreted. The two framings are complementary: TVOC combines the largest uncertainty ratio with a moderate d (≈0.8) because its within-kitchen variance is dominated by episodic cooking peaks, whereas the climatic channels show large d with modest uncertainty ratios. All effect sizes are descriptive; no inferential claim is attached (Sections 5 and 7).

Table 5. Kitchen-versus-indoor mean differences relative to propagated instrument uncertainty and descriptive effect size

6.2 Correlation structure

Figures 3 and 4 show the per-environment Pearson correlation matrices for the seven recorded channels. Both matrices are computed on all paired non-missing per-sample values (indoor n = 1,333; kitchen n = 1,253). These matrices are exploratory, as correlation does not establish causation. Furthermore, both matrices should be interpreted in light of the CCS811's on-chip algorithmic coupling: under quiet, low-VOC conditions, the chip's internal model drives eCO₂ and TVOC nearly deterministically from a common metal-oxide (MOX) baseline. Consequently, the very high indoor TVOC↔eCO₂ correlation reflects the on-chip algorithm rather than an atmospheric coupling. The indoor matrix accordingly shows a near-perfect TVOC↔eCO₂ correlation, together with moderate positive coupling between humidity and the MQ-series and a moderate inverse coupling between temperature and humidity.

3.png

4.png

The kitchen matrix shows the opposite pattern of interest: the TVOC↔eCO₂ correlation is lower than in the indoor matrix. This occurs because the chip's algorithmic coupling is partially broken by genuine, transient VOC injections from cooking, which the eCO₂ branch does not track in lockstep. Meanwhile, TVOC and eCO₂ both develop meaningful positive couplings with temperature and humidity, consistent with the expectation that cooking events simultaneously elevate heat, water vapour, and VOC emissions.

6.3 Cooking versus non-cooking sample separation

Under the cooking-window definition introduced in Section 5 (TVOC > 300 ppb at the per-sample level), 404 of the 1,253 kitchen samples (approximately 32%) qualify as cooking samples. Table 6 contrasts the cooking-window median with the non-cooking-window median for each kitchen channel. The TVOC and eCO₂ channels separate strongly between the two regimes (TVOC median ratio approximately 8×; eCO₂ median ratio approximately 2×), while the MQ-series and climatic channels show smaller separations.

Table 6. Kitchen samples separated by total volatile organic compound (TVOC) threshold

Across all kitchen samples, TVOC exceeded the 300 ppb threshold for a sample-weighted total of approximately 7.7 h (32% of the deployment) and kitchen eCO₂ exceeded 1,000 ppm—a commonly cited indoor-air-quality comfort threshold—for a sample-weighted total of approximately 15.3 h (64% of the deployment). Both figures are aggregations of all samples above the threshold (each kitchen sample weighted by the deployment-mean inter-sample interval of ≈68 s, i.e., the deployment duration of 23.80 h divided by the 1,253 samples); they are not the duration of any single contiguous cooking episode. The mean interval is used for this weighting rather than the 45 s median (Section 4.2) because the total sample count multiplied by the mean interval recovers the deployment duration, whereas the median does not reflect the longer inter-sample gaps caused by Wi-Fi reconnections and would understate elapsed time. Considered together with the calibration disclosure (Section 3.3) and the limitations (Section 7), these observations are consistent with the hypothesis that conventional exhaust ventilation may only partially mitigate cooking emissions at occupant distance under the observed operating conditions, but they are not a confirmatory population-level claim.

6.4 Hourly patterns and time-series visualisation

Figures 5-7 summarise the kitchen deployment at hourly and per-sample resolution. Figure 5 shows hourly-averaged TVOC and eCO₂, both peaking during the canonical lunch and dinner preparation windows.

5.png

Figure 5. Hourly-averaged kitchen total volatile organic compound (TVOC) and eCO₂ concentrations

6.png

Figure 6 shows the raw TVOC time series with cooking-window samples shaded, and Figure 7 shows the corresponding eCO₂ series, visually confirming that the TVOC and eCO₂ cooking peaks co-occur.

7.png

This pilot study presented a documented calibration and uncertainty framework for a low-cost air-quality monitoring stack (Arduino Uno + ESP8266 + MQ7 + MQ135 + CCS811 + GP2Y1010 + DHT11), with a transparent disclosure of the applied and unapplied conversions. The framework was applied to paired 24 h deployments in a regular indoor room and a hostel-mess kitchen, enabling a robust comparison of the two environments. Kitchen samples exhibited TVOC above 300 ppb for approximately 7.7 h (32%) and eCO₂ above 1,000 ppm for approximately 15.3 h (64%), with cooking vs. non-cooking sample medians separating strongly for TVOC and eCO₂. The dominant contrasts—eCO₂, TVOC, dust and humidity—exceeded the instrument-uncertainty bounds by factors of approximately 4–15, so the observed separations are not merely artefacts of sensor accuracy specifications. However, bounded by the calibration disclosure and the limitations discussed above—notably, the absence of co-located reference instruments and direct ventilation-operating-state measurements—these findings are consistent with the hypothesis that conventional exhaust ventilation may only partially mitigate cooking emissions at occupant distance under the observed conditions. They do not constitute a confirmatory population-level claim. Future work should involve multi-day, multi-kitchen deployments with reference co-location.

Four follow-on directions are proposed: first, instantiating the calibration framework in firmware (populating CCS811 ENV_DATA with DHT11 measurements, performing an in-situ R₀ calibration for the MQ-series, and measuring the GP2Y1010 dark-current V_oc offset); second, co-location with reference instrumentation (reference TVOC analyser, NDIR CO₂ sensor, and PM monitor); third, a controlled exhaust on/off comparison with air-change-rate measurement, to convert the present hypothesis-generating claim into a confirmatory ventilation-adequacy test; and fourth, multi-day, multi-kitchen deployments with contemporaneous cooking-event annotation.