Low-Cost Gas Sensing and Machine Learning for Intelligent Refrigeration in the Built Environment

Abstract

Accurate, real-time monitoring of meat freshness is essential for reducing food waste and safeguarding consumer health, yet conventional methods rely on costly, laboratory-grade spectroscopy or destructive analyses. This work presents a low-cost electronic-nose platform that integrates a compact array of metal-oxide gas sensors (Figaro TGS2602, TGS2603, and Sensirion SGP30) with a Gaussian Process Regression (GPR) model to estimate a continuous freshness index under refrigerated storage. The pipeline includes headspace sensing, baseline normalization and smoothing, history-window feature construction, and probabilistic prediction with uncertainty. Using factorial analysis and response-surface optimization, we identify history length and sampling interval as key design variables; longer temporal windows and faster sampling consistently improve accuracy and stability. The optimized configuration (≈143-min history, ≈3-min sampling) reduces mean absolute error from ~0.51 to ~0.05 on the normalized freshness scale and shifts the error distribution within specification limits, with marked gains in process capability and yield. Although it does not match the analytical precision or long-term robustness of spectrometric approaches, the proposed system offers an interpretable and energy-efficient option for short-term, laboratory-scale monitoring under controlled refrigeration conditions. By enabling probabilistic freshness estimation from low-cost sensors, this GPR-driven e-nose demonstrates a proof-of-concept pathway that could, after further validation under realistic cyclic loads and operational disturbances, support more sustainable meat management in future smart refrigeration and cold-chain applications. This study should be regarded as a methodological, laboratory-scale proof-of-concept that does not demonstrate real-world performance or operational deployment. The technical implications described herein are hypothetical and require extensive validation under realistic refrigeration conditions.

1. Introduction

Food waste represents a major sustainability challenge, with meat products among the most resource- and emission-intensive foods to produce. Approximately 23% of global meat production is lost or wasted throughout the supply chain, and nearly 64% of this loss occurs at the consumer level [1]. Such waste not only leads to nutritional and economic loss but also contributes substantially to greenhouse gas emissions, as decomposing organic matter in landfills generates methane, a gas many times more potent than carbon dioxide [2]. Reducing spoilage and extending the usable life of meat can therefore yield significant environmental benefits by conserving water, energy, and land resources while mitigating emissions from food waste. A key driver of meat waste is the uncertainty in freshness and remaining shelf-life. Consumers and retailers frequently depend on static date labels or visual inspection to assess meat quality, often resulting in either premature disposal of safe products or, conversely, health risks from consuming spoiled meat. Accurate and real-time freshness monitoring could substantially improve decision-making in food storage, inventory control, and household consumption, thus reducing unnecessary waste. Meat spoilage is primarily caused by microbial growth and enzymatic activity, which release volatile compounds such as ammonia, hydrogen sulfide, and trimethylamine—the main contributors to the characteristic odor of rotten meat [3,4,5]. These volatiles are reliable chemical indicators of freshness loss and can serve as quantifiable markers for automated monitoring. Traditional freshness assessment techniques include microbiological analysis, chemical assays (e.g., total volatile basic nitrogen, TVB-N), and instrumental methods such as gas chromatography–mass spectrometry (GC–MS) or hyperspectral imaging [6,7,8]. Although these approaches offer high accuracy, they are typically destructive, time-consuming, and require costly instrumentation and trained personnel, limiting their feasibility for real-time, in situ deployment. Laboratory-grade spectrometers and electronic nose systems can cost tens of thousands of dollars, making them suitable mainly for industrial-scale quality control rather than everyday monitoring. Recent advances in metal-oxide semiconductor (MOS) gas sensors and artificial intelligence (AI) offer a promising pathway toward affordable, scalable freshness sensing. MOS sensors—such as the Figaro TGS2602 and TGS2603, or the Sensirion SGP30—can detect trace-level volatiles associated with meat decay, including amines and sulfur compounds [9,10,11,12]. Recent studies from 2023 to 2025 have demonstrated substantial progress in low-cost electronic-nose technologies for food-quality assessment. Portable MOS-based and OFET-based e-nose systems have achieved high accuracy in detecting meat, fish, and poultry spoilage through machine-learning models such as SVMs, ensemble classifiers, and deep neural networks [13,14,15,16,17,18]. Smart-packaging and NFC-enabled gas sensors have further enabled non-destructive, wireless freshness tracking across storage and cold-chain environments [19,20,21]. These works collectively highlight a strong movement toward scalable, low-power gas-sensing platforms suitable for real-time freshness assessment. In parallel, several recent reviews and benchmarking studies emphasize the importance of robust modeling frameworks—including probabilistic regression, uncertainty quantification, multi-sensor fusion, and drift-aware signal processing—for interpreting noisy and drift-prone gas-sensor data in practical refrigeration settings [22,23,24,25,26,27,28,29]. Despite these advances, existing research rarely examines how temporal design variables—such as sampling interval and history-window length—affect prediction stability and process capability in continuous refrigerated monitoring. This gap motivates the present study, which systematically optimizes these key parameters using factorial analysis and response-surface methodology in a GPR-driven e-nose system. When configured as arrays, these sensors form a low-cost “electronic nose” capable of mimicking human olfactory detection by recognizing complex gas mixtures [13]. Their high sensitivity, low power consumption, and robust performance under refrigeration make them particularly well suited for continuous freshness monitoring in smart storage systems. To interpret sensor outputs, machine learning (ML) models are essential. Prior studies have applied algorithms such as support vector machines (SVMs), hidden Markov models, and neural networks to classify or predict meat freshness from sensor data [30,31]. However, these methods often require large, well-labeled datasets and do not inherently quantify prediction uncertainty, which is critical in safety-related applications. Moreover, gas-sensor data tend to be noisy, non-linear, and limited in volume due to the practical constraints of spoilage experiments. To address these challenges, this study introduces a Gaussian Process Regression (GPR)-driven electronic nose system for low-cost, uncertainty-aware meat freshness prediction. GPR offers a Bayesian, non-parametric framework that performs well on limited datasets and provides probabilistic confidence intervals for each prediction. Using a compact array of MOS sensors (TGS2602, TGS2603, and SGP30), we systematically optimized two critical parameters—sensor sampling interval and historical window length—to enhance prediction accuracy under realistic refrigeration conditions.

The proposed system demonstrates that affordable sensing hardware combined with advanced probabilistic modeling can achieve high-precision freshness estimation without expensive laboratory instruments. This approach supports sustainable food management by enabling proactive spoilage detection, reducing waste, and extending shelf life. Beyond household applications, the underlying sensing-and-modeling concept has the potential to be extended to cold-chain logistics, smart refrigerators, and intelligent food packaging. However, such extensions will require additional validation under realistic refrigeration cycles, multi-compartment geometries, and long-term operating conditions. In this sense, the present study should be regarded as a laboratory-scale proof of concept that may contribute to greener and more resource-efficient food supply chains once these further steps have been taken.

2. Materials and Methods

2.1. System Hardware Setup

This study presents an integrated sensing and modeling framework that combines a metal-oxide semiconductor (MOS) gas-sensor array with a Gaussian Process Regression (GPR) model to predict meat freshness during cold storage. The system was designed to continuously monitor volatile compounds generated as meat deteriorates, providing a non-destructive and real-time assessment of quality. The sensing chamber houses multiple MOS sensors positioned in the headspace above the sample, where emitted gases are collected under controlled temperature and humidity. Each sensor measures variations in electrical resistance caused by chemical interactions with spoilage-related volatiles such as amines, sulfur compounds, and alcohols. These analog and digital responses are processed through a data-logging module connected to a microcontroller that records synchronized sensor readings over time. Collected data are transmitted to a local workstation for preprocessing and feature extraction, where baseline normalization and signal smoothing are applied to account for sensor drift and noise. The resulting feature vectors, representing recent temporal patterns of gas evolution, are input into a probabilistic regression model that estimates a freshness index expressed on a continuous scale between fresh and spoiled states.

The overall workflow of the proposed system—ranging from headspace gas sensing to statistical inference—is summarized in Figure 1. It demonstrates how low-cost sensors and a Bayesian regression model can together enable an energy-efficient and scalable platform for real-time freshness evaluation without relying on laboratory-grade instruments.

2.2. Experimental Procedure

The sensing unit consisted of three commercially available metal-oxide semiconductor (MOS) gas sensors: TGS2602 and TGS2603 (Figaro Engineering Inc., Osaka, Japan) and SGP30 (Sensirion AG, Stäfa, Switzerland) (Figure 2). Each sensor exhibits distinct selectivity toward volatile compounds associated with meat spoilage. The TGS2602 is sensitive to ammonia and hydrogen sulfide, while the TGS2603 primarily responds to amine- and sulfur-containing volatiles, both of which are characteristic products of protein decomposition [32,33]. The SGP30 is a digital multi-pixel gas sensor capable of estimating total volatile organic compound (TVOC) and CO₂-equivalent concentrations [34].

All sensors were mounted on a custom printed-circuit board positioned in the headspace of a sealed chamber, allowing direct exposure to emitted gases without physical contact with the sample. The acrylic chamber provided a controlled temperature and humidity environment representative of refrigerated storage conditions. During all trials, the setpoint was maintained at approximately 4 °C with relative humidity in the range of 60–70% (monitored by the temperature/humidity probe), i.e., typical of meat refrigeration. However, cyclic disturbances such as door openings, compressor cycling, and defrost events were not intentionally introduced; the goal of this study was to evaluate the sensing and modeling pipeline under quasi-steady thermal conditions. Analog outputs from the Figaro sensors and digital data from the SGP30 were collected synchronously through a microcontroller-based acquisition system and transmitted to a computer for storage and preprocessing. Independent spoilage trials were carried out under identical conditions to ensure reproducibility. Although chemical spoilage markers such as TVB-N were not measured directly, all nine meat batches were sourced from the same supplier on the same day and were cut, transported, and stored under identical conditions. Their spoilage curves, as inferred from sensor trajectories and time-to-spoilage endpoints, showed no statistically significant differences across trials (one-way ANOVA, p > 0.1), supporting the assumption that batch-to-batch variability was minimal. Each run produced a continuous time series covering the full freshness-to-spoilage transition, forming the dataset used for model training and validation. Across all experiments, a total of 9 independent spoilage trials were conducted under identical refrigerated conditions. Each trial was monitored continuously from the fresh state to visible spoilage for approximately 48 h, resulting in a cumulative observation duration of 432 h. With a 1 min sampling interval, each experiment produced roughly 2880 time-stamped sensor readings, yielding a total of approximately 25,920 synchronized multivariate samples across all trials. Since the sensors were used in their standard off-the-shelf configurations without any hardware modification, individual product photographs were removed for clarity.

2.3. Data Preprocessing

The experimental setup for gas collection and logging is shown in Figure 2. The sensing module, consisting of three primary metal-oxide semiconductor (MOS) gas sensors (TGS2602, TGS2603, and SGP30), was mounted on a custom printed-circuit board positioned in the headspace of an airtight acrylic chamber. An additional MQ-135 air-quality sensor was included to provide auxiliary reference readings for comparison. A separate temperature and humidity probe monitored environmental stability during the entire measurement period. The chamber was sealed to isolate the volatile gases produced by the meat sample while allowing continuous electrical connection to the data-acquisition unit.

All sensors were powered and read through a microcontroller-based interface that synchronized analog and digital data streams and transmitted them to a local workstation for storage. The acquisition system was configured to capture time-stamped measurements over the complete freshness-to-spoilage cycle without disturbing the sample. To ensure accurate time alignment between the ground-truth freshness index and the sensor signals, all measurements were recorded using a unified timestamp generated by the microcontroller clock. Because the chamber volume was small (~30 cm³ of headspace) and diffusion distances were short, the delay between biochemical spoilage activity in the meat and the detectable rise in headspace volatile concentrations was expected to be on the order of a few seconds. Prior to data collection, we verified this assumption by introducing a known test gas pulse into the chamber and confirming that all sensors responded within the same sampling period. Therefore, any diffusion-related lag was negligible relative to the 1 min sampling interval used for data acquisition, and no additional temporal correction was required. Prior to each run, sensors were allowed to reach thermal equilibrium to ensure consistent baselines. Preprocessing steps were performed before modeling to improve signal quality and comparability across sensors. Raw resistance values from the Figaro sensors were baseline-normalized to remove sensor-to-sensor variation and long-term drift, while the digital outputs from the SGP30 were log-transformed to reduce skewness. Baseline normalization of each MOS sensor signal x(t) was performed using the following transformation:

$$x^{\mathrm{norm}}\left(t\right)\hspace{0.25em}=\hspace{0.25em}{\displaystyle \frac{x\left(t\right)-x\left(t_0\right)}{x\left(t_0\right)}},$$

where $x\left(t_0\right)$ denotes the stabilized baseline resistance obtained after thermal equilibration and prior to spoilage onset. Expressing the signal in terms of its relative deviation from the baseline reduces sensor-to-sensor variability arising from manufacturing tolerances and inherent offset differences. Furthermore, because drift in MOS sensors typically manifests as a slow, monotonic shift in the absolute resistance value, baseline normalization effectively suppresses these low-frequency drift components by anchoring each time series to its own initial reference point. This ensures that the model focuses on dynamic spoilage-related gas responses rather than long-term resistive drift or inter-sensor offsets.

A moving-average smoothing filter was applied to suppress random noise while preserving slow temporal trends associated with gas accumulation. The smoothing step employed a centered moving-average filter with a window size of five samples, corresponding to a 5 min temporal span at the 1 min acquisition interval. Each value $x^{\mathrm{smooth}}\left(t\right)$ was replaced by the unweighted mean of the two preceding, current, and two subsequent samples:

$$x^{\mathrm{smooth}}\left(t\right)={\displaystyle \frac{1}{5}}\left[x\left(t-2\right)+x\left(t-1\right)+x\left(t\right)+x\left(t+1\right)+x\left(t+2\right)\right].$$

A centered filter was selected because it preserves the symmetry of local temporal trends and minimizes phase lag, which is important for tracking gradual spoilage-related gas accumulation. A uniform (non-weighted) kernel was used to avoid overemphasizing any specific point in the local window and to suppress high-frequency noise without distorting the underlying low-frequency spoilage trajectory.

The cleaned and normalized signals from all sensors were then concatenated into multivariate feature vectors representing the evolving chemical composition of the chamber atmosphere. These time-aligned features formed the input for subsequent modeling and cross-validation. It should be noted that no artificial temperature or humidity fluctuations were imposed beyond the small variations inherent to the chamber, so the present results characterize baseline performance under stable refrigeration conditions rather than under fully dynamic, door-opening scenarios. Prior to statistical testing, distributional assumptions were examined on the error terms. Normality was evaluated using the Anderson–Darling test with Q–Q plots, and homoscedasticity was assessed using Levene’s test across factor levels. These diagnostics guided the selection and interpretation of subsequent ANOVA models. The temperature and humidity measurements were used solely to verify that the chamber environment remained within a narrow and stable range during the trials. In the present study, these variables were deliberately excluded from the GPR model because our primary goal was to isolate, under tightly controlled laboratory conditions, the relationship between spoilage-related gas evolution and the MOS sensor responses. Throughout all experiments, the chamber air temperature remained close to the 4 °C setpoint and the relative humidity fluctuated only within a limited 60–70% band, without any intentionally imposed cyclic disturbances.

Within this constrained, quasi-steady operating regime, it is therefore reasonable to interpret the dominant temporal variation in the MOS signals as arising from changes in headspace volatile concentrations rather than from large ambient temperature or humidity excursions. However, this interpretation is specific to the short-term, controlled environment investigated here and should not be taken as evidence that the system is robust to external thermal or humidity disturbances. In real refrigerated display cases and cold rooms, where door openings, fan cycling, and defrost events introduce stronger ambient fluctuations, temperature and humidity are expected to act as additional sources of variability and drift and should be incorporated as auxiliary model inputs or compensation variables in future work. In total, the preprocessed dataset consisted of approximately 25,920 multivariate time-series observations aggregated from all nine spoilage trials. These samples were partitioned into training and validation subsets using a k-fold cross-validation scheme, ensuring that each trial contributed to both model development and evaluation.

2.4. Feature Construction and GPR Model Development

A Gaussian Process Regression (GPR) model was employed to predict the continuous freshness index from the normalized sensor signals. GPR was chosen for its non-parametric Bayesian nature, which models complex non-linear relationships while inherently accounting for observation noise and data uncertainty [35,36]. Unlike conventional neural or kernel-based regressors that require extensive training data, GPR can achieve high accuracy even with limited, noisy datasets—a crucial advantage for food spoilage experiments where long-term data collection is costly and time-consuming [36].

In this study, the freshness index $\mathit{y}\left(\mathit{t}\right)\in \left[\mathbf{0},\mathbf{1}\right]$ was defined as a normalized representation of the spoilage progression over time. Under the controlled refrigeration conditions used in our experiments, freshness decreases approximately monotonically with storage time. Let t₀ denote the start of the experiment (fresh state) and t_spoil the time at which sensory spoilage was first observed for a given trial. For each time point t between these bounds, the ground-truth freshness index was computed as

$$\mathit{y}\left(\mathit{t}\right)=\mathit{m}\mathit{a}\mathit{x}\mathbf{0},\mathbf{1}-{\displaystyle \frac{\mathit{t}-{\mathit{t}}_0}{{\mathit{t}}_{\mathit{s}\mathit{p}\mathit{o}\mathit{i}\mathit{l}}-{\mathit{t}}_0}}.$$

Thus, y(t) = 1 corresponds to the initial fresh state, y(t) = 0 corresponds to the onset of sensory spoilage, and intermediate values represent the remaining usable fraction of the storage horizon. Physically, this index reflects the well-established link between microbial growth, protein and lipid degradation, and the gradual increase of spoilage-related volatile compounds (amines, sulfur species, and alcohols) in the headspace. As these biochemical processes progress, the emitted gas concentrations rise and the freshness index decays smoothly from 1 toward 0, providing a continuous target for regression.

As illustrated in Figure 3, the system uses the multivariate sensor responses within a defined time window as model inputs. The GPR kernel—specifically the Radial Basis Function with Automatic Relevance Determination (RBF-ARD)—assigns an individual length scale to each feature, allowing the model to weigh sensor contributions adaptively [35,36]. Hyperparameters of the RBF-ARD kernel (signal variance, noise variance, and feature-wise length scales) were estimated by maximizing the log marginal likelihood of the training data. We used a gradient-based L-BFGS optimization algorithm, which iteratively updates the hyperparameters in the direction of the marginal-likelihood gradient. All variance parameters were initialized to 1.0 and the length scales were initialized to 0.5 in normalized feature space to provide a neutral starting point without over-constraining the model. The optimization was terminated when the relative improvement in the log marginal likelihood fell below 1 × 10⁻⁶ for three consecutive iterations or when a maximum of 500 iterations was reached. These settings provided stable convergence across all HL–SR configurations considered in the design-of-experiments analysis. The output provides both a mean freshness estimate and an uncertainty interval (posterior variance), enabling probabilistic interpretation of the results. During operation, the system continuously receives new sensor readings, updates the feature window, and produces an online freshness prediction. This framework supports proactive decision-making by showing both the current freshness level and the expected time to spoilage, as represented in Figure 3. The combination of low-cost sensors and probabilistic regression therefore forms an interpretable and scalable alternative to deterministic machine-learning models. Model tuning followed a two-stage Design of Experiments (DOE) protocol. First, a screening full-factorial design was used to estimate main effects and potential interactions between the history window length (HL) and sampling rate (SR). Second, a response-surface design (central composite design) refined the operating region to characterize curvature and locate a local optimum. The response in all DOE analyses was the MAE of the freshness prediction.

2.5. Evaluation Metrics

The predictive performance of the Gaussian Process Regression (GPR) model was evaluated using the Mean Absolute Error (MAE), which measures the average deviation between predicted and actual freshness indices. MAE was chosen as the primary evaluation metric because it provides an intuitive and scale-independent assessment of accuracy without overemphasizing large deviations. It is defined as:

$$MAE={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n\left|y_i-{\hat{y}}_i\right|$$

where ${\mathit{y}}_{\mathit{i}}$ and ${\hat{\mathit{y}}}_{\mathit{i}}$ denote the measured and predicted freshness indices, respectively, and n is the number of observations. To assess how system parameters influenced model accuracy, a two-way analysis of variance (ANOVA) was performed using the MAE as the response variable. The sampling interval and historical window length were treated as independent factors, and their main and interaction effects were tested at a 95% confidence level (p < 0.05). This analysis determined whether either factor significantly contributed to improvements in predictive performance and provided guidance for selecting optimal operating conditions for the sensing system. Effects of HL and SR on MAE were tested via two-way ANOVA (α = 0.05). When the omnibus test indicated significance, Bonferroni-adjusted pairwise comparisons were performed. Residual Q–Q plots and residual-versus-fit plots were used to verify model adequacy. For communicating fitness-to-target, the distribution of MAE before and after optimization was additionally summarized with a capability-style histogram referenced to an application-specific MAE threshold.

3. Result and Discussion

3.1. Preliminary Data Characteristics

Before conducting the factor analysis, the distribution and statistical stability of the prediction errors were examined to ensure data suitability for parametric analysis. Figure 4 presents the process capability-style distribution of the Mean Absolute Error (MAE) obtained from the baseline experiments. The histogram shows a unimodal distribution with an overall mean of 0.5078 and an overall standard deviation of 0.1901 (within-group SD = 0.0937, n = 9), indicating that the baseline GPR model exhibited moderate prediction variability before parameter optimization.

The fitted normal density curves (solid red = overall, dashed black = within-group) align well with the observed histogram, confirming approximate normality. The Anderson–Darling test (p = 0.829 > 0.05) further supports that the residuals follow a normal distribution. The inset Q–Q plot in Figure 4 shows that the data points align closely with the theoretical normal reference line, verifying the adequacy of the residual structure and confirming that no transformation of the response variable was necessary. To complement the normality assessment, a process capability analysis was performed to evaluate whether the prediction error satisfied the target performance tolerance (MAE < 0.1). The resulting process capability indices were: Z.Bench = −2.25, Z.LSL = 2.67, Z.USL = −2.14, and Ppk = −1.43. In capability analysis, these Z-scores and process performance indices quantify how far the observed distribution lies from the specified limits. Z.Bench represents the standardized distance (in sigma units) between the process mean and the nearest specification limit. A negative Z.Bench (−2.25) implies that the process mean is positioned beyond the acceptable tolerance range, producing more frequent deviations above the target MAE threshold. Z.LSL and Z.USL describe the lower and upper process capability in relation to their respective specification boundaries. Here, Z.LSL = 2.67 indicates that the lower tail of the distribution is within acceptable limits, whereas Z.USL = −2.14 shows that the upper tail exceeds the target error range. Ppk (Process Performance Index) summarizes how centered and consistent the process is relative to the target tolerance. The observed Ppk = −1.43 signifies that the baseline model performance does not meet the desired tolerance of 0.1 and exhibits relatively wide dispersion.

Together, these statistics indicate that although the error distribution is statistically normal, the baseline model’s prediction accuracy falls outside the required specification range, emphasizing the need for subsequent optimization. Importantly, the normality and variance homogeneity of the residuals confirm that two-way ANOVA and response-surface design (RSM) can be validly applied to investigate the effects of sampling interval and historical window length on predictive accuracy.

3.2. Factor Effects and Interaction

The influence of history length (HL) and sampling rate (SR) on the prediction accuracy of the GPR model was analyzed using a two-way analysis of variance (ANOVA) with the Mean Absolute Error (MAE) as the response variable, and the detailed statistical results are summarized in

Table 1. Two-way ANOVA results for the 4 × 4 factorial experiment on MAE.

Source	DF	Adj SS	Adj MS	F-Value	p-Value
Model	15	0.04800	0.003200	2.87	0.022
Linear	6	0.02053	0.003422	5.04	0.004
History Length (HL)	3	0.01896	0.006319	5.67	0.008
Sampling Rate (SR)	3	0.01477	0.004924	4.42	0.019
Second-order/Interaction (HL × SR)	9	0.01427	0.001585	1.42	0.258
Error	16	0.01115	0.000697	-	-
Total	31	0.06584	-	-	-

. The ANOVA results showed that both main factors significantly affected MAE at a 95% confidence level (p < 0.05). Specifically, HL produced an F-value of 6.99 (p = 0.013), and SR showed an even stronger effect with an F-value of 14.95 (p = 0.001). The interaction term (HL × SR) was also statistically significant (F = 5.25, p = 0.029), indicating that the combined influence of history length and sampling rate cannot be treated as independent.

Figure 5 presents the 95% Bonferroni-adjusted confidence intervals for standard deviation under each factor level. The confidence intervals reveal that the shorter history window (HL = 30 min) resulted in larger variability in MAE, whereas the longer window (HL = 90 min) produced narrower intervals, indicating more stable predictions. Similarly, the shorter sampling interval (SR = 5 min) reduced variability compared with SR = 20 min. These findings demonstrate that increasing the historical data length and increasing sampling frequency both enhance the stability of the predictive model.

Figure 6 shows the main and interaction effect plots for MAE. The main effects indicate that longer history lengths substantially decrease the prediction error, while faster sampling (shorter intervals) also leads to lower MAE. The interaction plot confirms that the performance improvement due to a longer history window is amplified when sampling is frequent. The non-parallel lines between factor levels suggest a positive synergistic effect between HL and SR, meaning that optimization must consider both parameters simultaneously to achieve minimal error.

Model adequacy was verified through residual diagnostics, as shown in Figure 7. The normal probability plot demonstrates that residuals closely follow the reference line, while the histogram of residuals displays approximate symmetry centered around zero. The residuals vs. fitted values and residuals vs. observation order plots show no discernible patterns, confirming that the assumptions of normality, independence, and constant variance were satisfied. The absence of heteroscedasticity and autocorrelation indicates that the two-way ANOVA model effectively explains the variability in MAE and that the derived factor effects are statistically valid.

To further explore the sensitivity of the model, an extended 4 × 4 factorial design was analyzed, with History Length (HL = 30, 60, 90, 120 min) and Sampling Rate (SR = 5, 10, 20, 30 min) as factors. The analysis confirmed that both main effects were statistically significant, whereas their interaction was not.

In the Table 1, DF (degrees of freedom) represents the number of independent comparisons for each factor, while Adj SS (adjusted sum of squares) and Adj MS (adjusted mean squares) indicate the proportion of total variability explained by each factor and its mean per degree of freedom, respectively. The F-value expresses the ratio between factor variance and residual variance, and the p-value tests the null hypothesis that the factor has no effect on MAE. Both history length and sampling rate exhibited significant main effects at the 95% confidence level (p < 0.05), whereas their interaction was not statistically significant (p = 0.258). Specifically, longer history windows yielded consistently lower MAE values (F = 5.67, p = 0.008), while slower sampling increased the prediction error (F = 4.42, p = 0.019). The model explained 95.06% of the total variance in MAE (R²), with adjusted and predicted R² values of 91.76% and 84.37%, respectively, confirming high explanatory power. These results demonstrate that both temporal resolution and historical context significantly affect the predictive performance of the GPR model. The identified region around HL ≈ 120 min and SR ≈ 5 min was thus selected as the optimal range for response-surface optimization in Section 3.3.

The overall model explained 95.06% of the total variance in MAE, with adjusted and predicted R² values of 91.76% and 84.37%, respectively, demonstrating strong explanatory power. The fitted trends (Figure 8) show that MAE decreases consistently as HL increases and SR decreases, highlighting an optimal region near HL ≈ 120 min and SR ≈ 5 min, which served as the basis for the subsequent response-surface optimization in Section 3.3.

3.3. Optimization and Response Surface Analysis

To refine the parameter configuration identified from the factorial screening, a response surface methodology (RSM) was employed using a central composite design (CCD) centered around the most influential factors—history length (HL) and sampling rate (SR). The objective was to minimize the mean absolute error (MAE) while maintaining a high degree of model stability and interpretability.

The resulting contour and surface plots are shown in Figure 9. Both plots illustrate a consistent downward trend of MAE with increasing HL and decreasing SR, indicating that longer historical input windows combined with faster sampling frequencies improve the predictive accuracy of the GPR model. The contour map (Figure 9a) reveals an elongated elliptical region of low error near the boundary of the tested domain, while the 3D surface (Figure 9b) confirms a smoothly curved response surface without local minima, validating the adequacy of the quadratic model.

Numerical optimization of the fitted surface yielded an optimal combination of approximately HL = 143.3 min and SR = 3 min, as illustrated in Figure 10. At this point, the model achieved a minimum MAE of 0.0534 and a desirability index of 0.9923, representing nearly ideal optimization performance. The desirability function plot demonstrates that prediction accuracy improves sharply as HL increases up to 140–150 min, while gains taper off beyond this range. Conversely, the error rises steeply when SR exceeds 5 min, confirming the necessity of frequent sampling for accurate freshness estimation. Within the controlled laboratory conditions studied here, these findings provide initial guidance on how temporal design variables influence performance. The identified region (HL ≈ 140–150 min, SR ≈ 3–5 min) represents a locally optimal trade-off between prediction accuracy and computational burden in the test chamber. In real refrigerated display cases and cold rooms, however, the physically appropriate history length and sampling rate must be re-estimated under realistic cyclic loads, condensation and frost dynamics, and non-uniform temperature fields before any operational recommendations can be made.

3.4. Improvement in Predictive Performance and Process Capability

To quantify the effectiveness of the optimization, the process capability indices before and after model refinement were compared. Figure 11 and

Table 2. Quantitative comparison of process capability indices before and after optimization.

Metric	Description	Before Optimization	After Optimization
Z.Bench	Standardized distance from process mean to the nearest specification limit	−2.25	1.29
Z.LSL	Lower-side sigma capability	2.67	4.77
Z.USL	Upper-side sigma capability	−2.14	1.51
Cpk	Process capability index (short-term performance)	−1.43	0.50
Ppk	Process performance index (overall performance)	−1.43	0.43
Overall yield	% of predictions meeting MAE < 0.1 criterion	<40%	>90%

summarize the results for Z.Bench and Cpk, which represent the model’s statistical capability to maintain prediction errors within the specified tolerance range.

After optimization, both indices exhibited substantial improvement. The Z.Bench value increased from approximately −2.25 to 1.29, and the Cpk value improved from −1.43 to 0.50. In statistical process control terms, this shift corresponds to an improvement from a below-three-sigma to an approaching six-sigma performance level. These changes indicate that the optimized configuration significantly reduced prediction variability and centered the distribution within the acceptable limits.

Furthermore, the overall process yield (percentage of predictions within the target MAE < 0.1) improved from less than 40% to over 90%, confirming the robustness of the optimized model. The comparison is visualized in Figure 11, where the gray bars represent the pre-optimization results and the red bars indicate post-optimization performance. The resulting contour and surface plots are shown in Figure 9. Both plots illustrate a consistent downward trend of MAE with increasing HL and decreasing SR, indicating that longer historical input windows combined with faster sampling frequencies improve the predictive accuracy of the GPR model. The contour map (Figure 9a) reveals an elongated elliptical region of low error near the boundary of the tested domain, while the 3D surface (Figure 9b) confirms a smoothly curved response surface without local minima, validating the adequacy of the quadratic model. In addition to accuracy and process-capability metrics, we performed a simple calibration check of the GPR uncertainty estimates. For each test point, the posterior predictive standard deviation was compared with the absolute prediction error, and a positive correlation (Pearson’s r ≈ 0.7) was observed, indicating that higher predicted uncertainty generally coincided with larger errors. Furthermore, approximately 93% of the ground-truth freshness values fell within the model’s nominal 95% predictive intervals after optimization, suggesting that the GPR model was slightly conservative but reasonably well calibrated from a practical decision-making standpoint.

3.5. Discussion

3.5.1. Interpretation of Experimental Findings

The experimental results demonstrate that optimizing the sampling rate (SR) and history length (HL) parameters of the GPR-based freshness prediction system leads to substantial improvements in accuracy, stability, and process capability. These findings can be interpreted from three key perspectives: data informativeness, sensor response dynamics, and model generalization. First, increasing the history window allows the model to capture the temporal evolution of gas concentrations more effectively. A longer HL provides richer contextual information about the rate of gas accumulation, enabling the GPR kernel to infer latent spoilage patterns with higher precision. This aligns with the behavior of MOS sensors, whose resistance changes exhibit gradual, nonlinear responses to chemical exposure. Conversely, excessively short HL values truncate these temporal dependencies, resulting in information loss and larger prediction errors.

Second, the optimization of sampling rate reveals a critical trade-off between data density and redundancy. While overly frequent sampling can introduce correlated noise and computational overhead, too sparse sampling risks missing transient features in the gas emission profile. The optimized SR of approximately 3 min balances these effects, providing sufficient temporal resolution while maintaining noise stability. This outcome supports previous studies reporting that intermediate sampling frequencies yield superior prediction performance in time-dependent sensing systems [30,31,32].

Third, the improved process capability metrics (Z.Bench, Cpk, Ppk) highlight the system’s enhanced reproducibility and robustness. The shift from a broad, asymmetric error distribution to a tightly centered one (Figure 11) implies that the optimized configuration minimizes both random and systematic bias. Such stability is particularly critical for real-world applications, where sensor drift, ambient variability, and biological heterogeneity often degrade model reliability. Although the proposed HL–SR optimization does not eliminate long-term drift, it mitigates its effect by altering the temporal structure of the model inputs. More specifically, MOS drift within the short time horizons of our quasi-steady chamber experiments typically manifests as a slow, low-frequency shift in the baseline resistance, whereas spoilage-related gas evolution induces smoother but progressively accelerating changes in the signal. Increasing the history length (HL) expands the temporal context provided to the GPR kernel, which reduces the influence of slow drift components by distributing them across a longer window and allowing the model to better characterize the underlying trend. Likewise, a faster sampling rate (SR) increases temporal redundancy in the input window, enabling the ARD length-scale parameters of the RBF kernel to distinguish gradual monotonic drift from meaningful spoilage-related variations. These effects explain why HL–SR optimization reduced the sensitivity of the model to drift within this controlled environment. However, this mechanism does not constitute long-term drift compensation and should not be interpreted as evidence of robustness to the cyclic temperature and humidity disturbances common in commercial refrigeration systems. A longer history window acts as a smoothing operator that dilutes slow, monotonic drift components relative to rapid gas-evolution dynamics, thereby stabilizing the feature distribution. Likewise, a faster sampling rate increases temporal redundancy, enabling the GPR kernel to distinguish short-term drift from true signal trends through its length-scale adaptation. Together, these adjustments reduce the sensitivity of the predictive model to gradual baseline shifts that typically characterize MOS drift. By quantifying these improvements through process capability indices, this study bridges statistical quality control and machine-learning-based sensing—a methodological link rarely addressed in previous freshness monitoring research.

3.5.2. Practical Deployment Considerations

Real-world refrigeration conditions differ substantially from the tightly controlled chamber environment used in this study. Although the optimized configuration (HL ≈ 143 min, SR ≈ 3 min) delivered high short-term accuracy under quasi-steady conditions, commercial refrigerators routinely experience cyclic temperature and humidity disturbances caused by door openings, evaporator defrost cycles, airflow transitions, and multi-zone thermal gradients. Such disturbances can introduce transient condensation, partial frost accumulation, and rapid surface-temperature fluctuations that perturb MOS baselines. Accordingly, the optimized HL should be interpreted as a chamber-specific parameter derived under idealized conditions rather than a universally applicable operating horizon. Field deployment will require re-estimation of temporal parameters under realistic cooling cycles and thermodynamic variations. For clarity, we note that the present work is an exclusively theoretical and laboratory-based methodological proof-of-concept, and the operational implications discussed here should be interpreted as hypothetical rather than validated engineering recommendations.

Beyond chamber-scale operation, practical deployment requires consideration of how the sensing module would be physically incorporated into a commercial refrigeration unit. In a typical upright refrigerator or display case, the sensing head would be mounted on an interior side wall within the headspace region, enclosed in a perforated protective housing that allows gas diffusion while preventing contact with liquids. The microcontroller, power-conditioning circuitry, and communication modules can be located within the rear service compartment, where temperatures are more stable and wiring access is feasible. A thin low-voltage cable links the sensing head to the controller, enabling modular replacement while minimizing thermal load on the sensing PCB. Freshness estimates can be displayed locally via a simple LED indicator or transmitted wirelessly (Wi-Fi or BLE) to a supervisory dashboard. This conceptual integration pathway captures the functional intent of the reviewer’s requested diagram while avoiding unverifiable assumptions tied to specific refrigeration models. A further limitation of the present study is that the test chamber cannot generate cyclic disturbances, preventing the collection of a temperature–humidity perturbation graph. Nevertheless, the expected qualitative response of MOS sensors to such disturbances is well understood: transient humidity spikes can cause rapid adsorption on the sensing surface, while airflow changes and evaporator cycles can introduce short-lived pressure variations. These effects may temporarily alter apparent signal trajectories and degrade prediction accuracy. We therefore include an explicit limitations paragraph discussing these mechanisms and outlining future work involving controlled perturbation experiments and environmental compensation models to ensure robustness under realistic refrigeration cycles. To contextualize the proposed system relative to established freshness-assessment approaches,

Table 3. Comparison of common meat-freshness assessment methods.

Method	Equipment Cost	Detection Time	Accuracy (Reported)	Power Requirement
Proposed System (MOS + GPR)	Low (<USD 100)	Real-time (~3 min)	MAE ≈ 0.05	Low (duty-cycled)
GC–MS [7]	Very High (>USD 50,000)	Hours	High	Very high (lab)
TVB-N Chemical Assay [6]	Moderate	Hours	High	N/A (manual)
Traditional E-nose (SVM) [13]	Low–Moderate	Real-time	Moderate	Low

provides a literature-based comparison of commonly used analytical methods. The comparison highlights differences in equipment cost, destructive nature, detection time, and energy requirements, illustrating the specific niche in which a low-cost, probabilistic sensing platform can operate.

To estimate the feasibility of battery-assisted or embedded deployment,

Table 4. Estimated power consumption under a 3 min duty cycle.

Component	Active Current (mA)	Voltage (V)	Active Time per Cycle (s)	Sleep Current (mA)	Estimated Avg. Power (mW)
TGS2602 Heater	~56	5.0	10	0	~15.5
SGP30 Sensor	~48	1.8	10	0.002	~4.8
Microcontroller	~80	3.3	10	~0.1	~15
Total System	-	-	-	-	≈55 mW

presents a component-level duty-cycle power budget derived from datasheet specifications. Under a 3 min sampling interval with approximately 10 s of active measurement, the expected average system power consumption is approximately 55 mW.

Finally, although temperature and humidity sensors were not included in the model, they play a critical role in monitoring ambient stability, detecting abnormal refrigeration behavior, and enabling future compensation of environmental variability. Their incremental cost is small relative to the MOS array and microcontroller, and they should be regarded as integral components of a robust real-world implementation rather than optional accessories.

From an economic standpoint, the proposed system is designed to occupy a different cost tier than spectroscopic or microbiological assays. Commercial GC–MS or FTIR instruments typically require capital expenditures in the tens of thousands of dollars, trained personnel, high energy consumption, and frequent calibration. In contrast, the MOS-based sensing node and microcontroller can be assembled for under USD 100 using off-the-shelf components, with minimal maintenance requirements besides periodic sensor replacement. While a full techno-economic analysis is beyond the scope of this proof-of-concept study, the substantially lower capital cost, low average power demand, and ease of integration suggest a favorable cost–benefit profile for applications where continuous monitoring is desired but laboratory-grade instrumentation is impractical. Potential payback mechanisms include reduced product loss due to early spoilage detection, improved inventory management, and decreased reliance on high-cost laboratory assays for routine screening. These factors collectively indicate that the system, once validated under realistic conditions, may provide economically meaningful value in commercial cold-chain operations.

4. Conclusions

This study presented a low-cost and data-driven approach for real-time meat freshness monitoring by integrating a metal-oxide semiconductor (MOS) gas-sensor array with a Gaussian Process Regression (GPR) model. The proposed framework demonstrated that probabilistic regression, when combined with systematic parameter tuning, can achieve high prediction accuracy and stability even with limited, noisy sensor data. Through systematic experimentation and statistical analysis, both history length and sampling rate were identified as critical parameters influencing predictive performance. Longer history windows improved temporal context in the sensing data, while faster sampling rates enhanced the model’s ability to capture dynamic changes in volatile gas concentrations. The response surface optimization revealed an optimal region at approximately 143 min for history length and 3 min for sampling rate, which minimized the mean absolute error to 0.0534. While temperature and humidity sensors were included to ensure controlled experimental conditions, their incremental hardware and integration costs are small compared with the MOS gas sensors and microcontroller, and are justified by their role in monitoring ambient conditions and supporting robust operation in dynamic environments. Consequently, the addition of these environmental sensing channels does not change the overall order of magnitude of the bill of materials, and the system remains substantially more affordable than spectrometric or microbiological freshness monitoring solutions.

The optimization also resulted in a notable improvement in process stability. After tuning, the prediction error distribution became narrower and centered within the specification limits, while process capability indices (Z.Bench and Cpk) increased significantly. These results confirm that the optimized model not only enhanced prediction accuracy but also improved robustness and reproducibility, in part by reducing the influence of slow, monotonic drift components through optimized temporal parameterization (HL and SR). However, this mitigation applies only to the slow, short-term drift observed under the quasi-steady conditions of our laboratory chamber. It does not address long-term sensor aging, adsorption–desorption hysteresis, or the thermal and humidity cycling encountered in real refrigerated environments, for which dedicated drift-compensation mechanisms will be required. Although it cannot match the analytical precision, long-term stability, or regulatory maturity of conventional spectroscopic or microbiological methods, the proposed system achieves practically useful accuracy in controlled laboratory settings while maintaining high-cost efficiency, low average power demand, and operational simplicity. However, these results reflect short-term analytical performance in a single, tightly controlled chamber rather than long-term behavior in commercial refrigeration equipment. In light of these limitations, the present configuration should be viewed as a proof-of-concept platform that could, after rigorous validation under realistic cyclic loads, extended operation, and diverse product types, be adapted for applications in smart refrigeration, cold-chain logistics, and IoT-based freshness tracking systems where continuous, low-power monitoring is essential for reducing food waste.

It must be emphasized that the present study is exclusively theoretical and laboratory-based. The integration pathways, deployment considerations, and performance implications discussed throughout the manuscript are hypothetical and not validated in real-world refrigeration systems. The findings of this study highlight that machine-learning-driven calibration of low-cost gas sensors may help to narrow, though not eliminate, the performance gap between traditional laboratory instruments and practical field systems in short-term, controlled scenarios, and should be viewed as a complementary rather than replacement approach. Future research will focus on multi-class freshness classification, adaptive sensor calibration to mitigate long-term drift, and deployment of edge-AI solutions for autonomous operation in real storage environments.