I. INTRODUCTION

The global rise in greenhouse gas concentrations has accelerated climate change, resulting in growing temperatures, sea-level rise, and ecological imbalance. Monitoring gases such as carbon dioxide (CO₂), methane (CH₄), and nitrogen oxide (N₂O) is essential for regulatory compliance, environmental modelling, and urban air-quality assessment ^[1]. Traditional sensing systems often rely on infrared spectroscopy or chromatography, which provide high accuracy but are expensive and power-intensive, making them unsuitable for large-scale deployment ^[2].

In contrast, semiconductor metal-oxide (MOS) gas sensors are low-cost, lightweight, and easily integrated with other systems ^[3]. In addition, it improves detection limits and offers more accurate calculations of volatile organic compounds enabling applications across several fields ^[4]. Despite these advantages, MOS sensors suffer from drawbacks such as non-selectivity, cross-sensitivity to humidity and temperature, and significant energy demand from heater elements. These challenges necessitate intelligent sensing mechanisms that can dynamically balance power consumption and improve accuracy ^[5].

Recent advancements in nanostructured materials have opened new avenues for improving gas-sensing performance. The specific design of 2D nanostructures, for example, nanosheets, has gathered attention for their ability to form a surface region and provide plenty of energetic sites for gas adsorption ^[6]. The morphology of nanosheets permits greater coverage of surface particles, thereby enhancing gas-sensing efficiency ^[7]. One of the challenges associated with SnO₂-based gas sensors is their high operating temperature, typically 200 to 400 °C. This temperature range may result in material deformation, particularly if the shape is not optimized for thermal stability. Due to their intrinsic stability, spherical nanoparticles are better suited to withstand these conditions. Their straightforward geometric design reduces agglomeration and provides reliable gas diffusion pathways, thereby improving sensor performance and reaction times. Because of these characteristics, spherical SnO₂ nanostructures are among the most often used options for gas sensing applications ^[8].

The current work builds on these ideas by concentrating on the synthesis and characterization of SnO₂ nanostructures produced utilizing a one-step hydrothermal procedure. The performance of gas sensors fabricated from the synthesized materials is assessed against greenhouse gases. The gas-sensing approach that underlies the materials' selective response is also examined. Through the use of spherical morphology's benefits, including increased stability, sensitivity, and selectivity, this research aims to determine current issues and advance the development of high-performance gas sensors ^[9].

Motivation of this research: The recognition and control of CO₂ radiation is due to its important influence on global warming. The global atmospheric concentration of CO₂ reached 423.6 ppm in 2024, a level of emissions in history. A SnO₂ nanostructure can be obtained by applying chemical systems containing metal oxides to improve the response of these sensors. Methane is a hydrocarbon, volatile, and neutral gas, but it is recognized as extremely flammable. As stated in the report of the National Oceanic and Atmospheric Administration (NOAA), atmospheric CH₄ concentration was measured at almost 1921.79 parts per billion (ppb) in 2024. Hence, the recognition of CH₄ is important owing to security concerns, though its detection is very difficult due to its non-polarity as well as the great enthalpy of the C-H bonds. Another significant gas, like N₂O, is a colorless gas with a sweet odor. The World Meteorological Organization stated that the N₂O concentration in the air was 337.7 ppb in 2024. N₂O detection is a significant step toward improving N₂O gas sensors, supporting public health and environmental safety.

Problem Statement: MOS sensors, such as SnO₂, zinc oxide (ZnO), and doped composites have been extensively investigated for the detection of CO₂ and CH₄ gases. While these approaches enhance sensitivity, classification across multiple greenhouse gases remains problematic because of overlapping response profiles ^[10]. Supervised Machine Learning (ML) algorithms, such as Support Vector Machine (SVM), artificial neural networks (ANN), Convolutional Neural Network (CNN), and k-nearest neighbor (K-NN) have been applied for gas categorization from sensors. Deep learning with transitory responses improves accuracy. The algorithm can be utilized with an MOS sensor for gas categorization, but it relies on static sample rules and does not adjust control sensor functions. This static sample rule creates suboptimal power productivity and inadequate flexibility ^[11].

RL has shown potential in adaptive sampling, power optimization, and feature attainment. But RL-based gas recognition is a sequential decision-making problem, balancing detection accuracy. RL is a type of ML in which an agent learns best rules via communication with its environment, provides a promising solution. By establishing gas detection as a consecutive decision-making problem, RL can strongly control heater temperature, sampling intervals, and sensor activation rules to improve accuracy.

This approach addresses the gap by proposing an RL-based adaptive gas-detection mechanism that influences the MOS sensor to detect greenhouse gases in the environment. This paper introduces an RL-based gas categorization system that utilizes MOS sensors for greenhouse gas recognition and categorization. The novelty rests in joining the MOS sensor with adaptive RL-driven rules that learn to classify gases like CO₂, CH₄, N₂O while optimizing power and reliability. The main objective of this paper is to enhance the dynamic properties of SnO₂ via surface modifications and nanostructures, thereby improving the greenhouse gas detection performance of MOS sensors.

The remaining tactic is assembled as follows. Section II reviews greenhouse gas detection using ML learning algorithms. SnO₂ nanostructure applied as gas sensors with RL for categorization of greenhouse gases is covered in Section III. The results and discussion are presented in Section IV. Lastly, Section V completes the proposed system.

II. RELATED WORKS

The MOS sensors are robust and inexpensive, but they require high power due to heating, have limited selectivity that confuses oxygen with other gases, and exhibit stability issues. However, nanostructures with micro-electro-mechanical systems (MEMS) integration also include ML algorithms for adjustment to provide faster response and enhance accuracy ^[12]. Supervised ML algorithms, such as SVMs, k-NN, and DNNs, have shown strong performance in categorizing gas types and estimating concentrations. However, these algorithms accept static data gathering and do not optimize sensor operation. Adaptive sensing methodologies have recently gained attention. Active feature-gaining and adaptive sampling methods save power but have not been clearly applied to greenhouse gas observation with MOS sensors ^[13]. ML algorithms, such as principal component analysis (PCA), SVM, KNN, and neural network have been extensively used with multiple sensors to improve efficiency. The drawbacks of these ML models increase costs and power consumption ^[14].

For disease diagnosis, semiconductor sensors are used to improve patient health. ML models applied to enhance the identification ability of a sensor that analytically lists, comprising SVM, random forests, ANN, and PCA. The study emphasizes the substantial potential for early illness identification and the complementary roles that gas sensor arrays and ML play in non-invasive medical diagnostics. However, challenges include sensor drift, inter-individual breath variability, limited datasets, and environmental factors that impede real-world clinical adoption ^[15]. Electronic nose (e-nose) technology, which uses chemoresistive sensors to specifically detect a target gas, has attracted significant interest for a range of applications, including personal health monitoring and smart factories ^[16]. Advanced gas sensors consider hydrogen sulphide (H₂S) detection crucial. The goal of the research is to make a substantial contribution to the creation of highly effective, reasonably priced gas sensors by examining the gas-sensing characteristics of nano-sized pure ZnO and the effects of different production techniques and ultrasonication treatment. The sensitivity of mixed-gas situations necessitates strict process control, which remains a probable risk ^[17].

Thermal imaging cameras and gas-detection sensors are used to simultaneously capture data samples in the multimodal gas data collection. Since low-cost sensors are often less dependable and sensitive, they can't detect gases at greater distances. To get over the limitation of relying only on sensors for gas detection, a thermal camera that can detect fluctuations in temperature is additionally utilized during the aggregation of the present multimodal dataset. The dataset is valuable but is constrained by inadequate gas sampling, controlled lab conditions, sensor requirements, and the absence of noise, which limit direct deployment ^[18]. A pattern recognition algorithm for a chemiresistive gas sensor focuses on selectivity and stability, and the use of cutting-edge technologies such as ML, nanotechnology, and hybrid materials has greatly enhanced sensor performance. Techniques, including pattern recognition, feature extraction, response signal analysis, and others, are often used together. These techniques effectively correct drift by more thoroughly analysing and processing sensor signals, significantly improving the handling and clarity of sensor data. These developments open new avenues for advancing and using gas detection technologies ^[19].

A single micro-LED (μLED)-embedded photoactivated (μLP) gas sensor employs time-variant illumination to identify the species and amounts of target gases. Forced transient sensor responses are produced by applying a fast-changing pseudorandom voltage input to the μLED. For gas recognition and valuation of concentration, the acquired complicated transient signals are analysed using a DNN. The cost, area, and power consumption of e-nose technology are all improved by this method ^[20]. A dual-mode gas sensor has been developed that uses pulsed-temperature operation to distinguish multiple single-component gases. The sensor has an integrated high efficiency microheater and SnO₂ nanosheets that provide enough reaction sites. Scalable manufacturing and device size reduction are made possible by the fabrication method's compatibility with common MEMS techniques. This mechanism successfully decouples the effects of temperature, diffusion properties, and reaction kinetics on sensing signals by repeatedly supplying the microheater with pulsed power inputs. Although selectivity is increased by a pulse-driven MEMS gas sensor with ML, ambient interference, sensor drift, reliance on dataset, constrained test conditions, and reproducibility issues prevent widespread use ^[21].

To increase safety and prevent mishaps caused by poisonous and explosive gases, semiconductor-based gas sensors are applied to detect hazardous gases. Advances in semiconductor sensors, like MOS sensors that integrate with MEMS for miniaturization also include wireless features for monitoring hazardous gases. It provides better sensitivity, selectivity, and low-temperature operation, and addresses concerns such as high-power requirements and false alarms caused by CH₄ interference. Broad adoption is still hampered by real-world issues such as high operating temperatures, cross-sensitivity, humidity-related issues, and long-term reliability ^[22]. A wireless system is used to detect combustible gases such as H₂ and CO using a sensor microcontroller with Zigbee communication. This system applies particle swarm optimization for overall tuning and the dung beetle optimization algorithm for local adjustments that enhance the gas detection accuracy ^[23]. Table 1 presents a comparison of related work.

III. PROPOSED METHOD

The proposed system provides a controllable gas environment. The gas classification model using the SVM algorithm efficiently categorizes CH₄ and CO gases using semiconductor sensors such as MQ4 and MQ7 ^[24]. Hence, the proposed system takes MQ4 and MQ7 semiconductor sensors to categorize greenhouse gases. Here, the MQ sensor is sensitive to CO₂, CH₄, N₂O, temperature, and humidity, and is combined with microcontrollers to obtain data. Then, apply the preprocessing module to normalize signals, remove temporal features, and perform noise reduction. In the proposed system, the RL agent selects actions such as heater temperature adjustment, sampling rate selection, and sensor activation to categorize the gases. In addition, the RL reward function balances detection accuracy, power efficiency and sampling delay. Fig. 1 illustrates the framework of the proposed system.

Fig. 1. Framework of the proposed system.

The proposed system framework consists of a gas chamber attached with gas cylinders, data acquisition equipment (DAQ) to control and measure gas sensors. A source measure unit (SMU) delivers a certain voltage, and a digital multimeter (DMM) verifies electrical signals in the circuit. In addition, the proposed framework utilizes a mass flow controller (MFC) for precise control of CO₂, CH₄, and N₂O flow, together with a computer to activate these components and execute gas categorization algorithms. Furthermore, Gas sensors generate electrical impulses by identifying the presence of specific gases. MQ gas sensors are the best option because they are small, fast responding, and long-lasting. Every sensor has a heating element that generates a gas concentration by providing a basic output voltage.

Here, MQ4, MQ7, MQ135, and DHT-11 sensors are used to distinguish CH₄, N₂O, CO₂, temperature, and humidity, respectively, as exposed in Fig. 1. These n-type semiconductor sensors are based on SnO₂ nanostructures and are triggered by interactions with flammable gases close to the SnO₂ surface. Oxygen is intensely absorbed on the sensor's surface during heating, acquiring electrons from the SnO₂ surface and creating an electron-reduction zone beneath it. When SnO₂ nanostructures are exposed to CH₄, N₂O, and CO₂ gases with adequate energy, these gases react by adsorbing oxygen atoms, which release electrons to SnO₂ and lower sensor resistance.

The data collection circuit must thus include a load resistor, and when the sensor resistance decreases due to gas interaction, the voltage drops across the load resistor increases. As a result, it gathers load resistor voltages for gas categorization. Fig. 2 demonstrates the block diagram of the greenhouse gas categorization function. The proposed system elements include data collection, feature extraction, a simulator, a design RL algorithm, and training and testing.

Fig. 2. Block diagram of the proposed system.

The data collection process obtains gas information from several SnO₂ nanostructure-based gas sensors that evaluate CO₂, N₂O, CH₄, temperature, and humidity in the atmosphere. The gas information contains noise from electromagnetic interference, signal communication failures, and variations in the experimental atmosphere. The function of data preprocessing is to remove noise and intervention data from the gathered signal. For training the identification method, as many key features as feasible that can signify the target gas must first be removed. The feature dataset derived from these initially extracted features is usually high-dimensional. Then remove any unnecessary information. Finally, to recognize the target gas, the enhanced feature dataset is utilized to train the gas identification model using an RL algorithm.

IV. EXPERIMENTAL ANALYSIS

Here, the proposed system showed the gas categorization investigation for a particular gas setting. More precisely, when SnO₂ nanostructure-based semiconductor gas sensors are used, a circumstance is considered to recognize if it presents CH₄, CO₂, and N₂O. Experiments with gas injection begin with a maturation period to heat the sensors, then a relief period. Gas is introduced at predetermined rates for 20 seconds after a sufficient maturation period. After the injection is complete, a response time is provided to ensure the gas sensors have fully reacted. Until the desired gas concentration is attained, the same procedure is repeated while increasing the gas levels.

Gas injection experiments initiated with maturing time to heat and, after this, alleviate the sensors. After adequate maturing time, gas is inserted at specific rates for 20 seconds. Injection was then completed, and the response time was allowed to ensure the gas sensors fully responded. The same process is repeated, raising gas levels until the target gas concentration is reached. The tests are conducted in an environment consisting of air for the categorization of CH₄, CO₂, and N₂O gases, since the metal oxide SnO₂ nanostructure semiconductor sensors cannot work without oxygen. After guiding every gas experiment, the CH₄, CO₂, and N₂O gases were introduced into the chamber to initialize the experimental environment.

The proposed system utilizes an RL algorithm for categorizing CO₂, CH₄, and N₂O gases using the experimental setup. In general, RL is an ML model categorizing two or more data classes. The dataset can be retrieved using https://gml.noaa.gov/data/dataset.php?item=cob-co2-aircraft-pfp. The proposed system evaluates metrics such as sensor output voltages (SOVs), gas response, accuracy, rate of false negative (RFN), and rate of false positive (RFP). The confusion matrix (CM) is a categorization performance. New datasets, not part of the training data, were used to verify the model's reliability using the CM. The matches between the forecasted and actual classes are shown in the CM. Fig. 3 shows the CM for greenhouse gas categorization.

Fig. 3. CM for greenhouse gases categorization.

1. Analysis of greenhouse gas response

The gas response refers to the assessable change in a sensing system output triggered by the concentration, presence, or variation of greenhouse gases, such as CO₂, CH₄, and N₂O. The gas response computation is specified below.

(2)

$\text{Gas response} = (R_{air} - R_{gas})/R_{air}$

$R_{air}$ indicates the sensor resistance in fresh air, and $R_{gas}$ represents the resistance in the presence of CO₂, CH₄, and N₂O. This gas response shows how powerfully the sensor responds to greenhouse gases.

The operation of greenhouse gas-sensitive materials relies heavily on their optimal operating temperature. To analyze this feature, the SnO₂ nanostructures' gas sensitivity is examined for a variety of gases at varying working temperatures, such as CH₄, CO₂, and N₂O. The findings show a distinctive rock-shaped temperature-response curve for SnO₂ nanostructure, as shown in Fig. 4.

Fig. 4. Greenhouse gas response of the SnO₂ nanostructures towards 500 ppm greenhouse gases at the temperature ranges between 100 to 400 °C.

This behaviour is explained by the conflict between structural distortions or reduced greenhouse gas adsorption effectiveness at higher temperatures and enhanced surface activity at intermediate temperatures. The gas reaction gradually improves as the temperature rises, provided it is run below 250 °C. The increased activity of surface-adsorbed molecules, which permits more effective electron exchange between the material and the greenhouse gas, is primarily responsible for this improvement.

Because of the restricted mobility of charge carriers at lower temperatures, the sensing response is suppressed. On the other hand, the material undergoes structural distortions that weaken its surface integrity when temperatures exceed 250 °C. Furthermore, excessive heat makes the target greenhouse gas molecules too mobile, shortening their residence time on the sensor surface and preventing effective interaction between the material and the sensor. Amongst the verified greenhouse gases, N₂O illustrated the greatest response compared to CO₂ and CH₄. Conversely, the CH₄ gas displayed weaker responses, with CO₂ showing the smallest sensitivity. Thus, the optimal functioning temperature for the SnO₂ material was determined to be 250 °C that is utilized for all succeeding performance evaluations.

The results demonstrate a significant improvement in responses to gases such as CO₂, N₂O, and CH₄. SnO₂ nanostructure-based sensor by the mixing method and described a lesser response time (less than 20 s) at 250 °C. The results denote that the response time of CH₄ is 20 sec at 250 °C. Compared with the nanostructure, the mat-like nanostructure shows greater sensitivity, with a response time of 25 sec. Fig. 5 demonstrates the greenhouse Gas response of SnO₂ nanostructures towards several concentrations of gases from 50 to 500 ppm at 250 °C.

Fig. 5. Greenhouse gas response of SnO₂ nanostructures towards several concentrations of gases from 50 to 500 ppm at 250 °C.

The response of gas sensors is also closely related to the gas concentration. To explore this connection, the gas response of the SnO₂ nanostructures material is observed for concentrations fluctuating from 50 to 500 ppm at the optimum functional temperature of 250 °C. As shown in Fig. 5, the signal sensitivity improves steadily with increasing gas concentration for all established gases. Significantly, the response did not plateau over the tested concentration range, demonstrating that the SnO₂ nanostructures possess sufficient reaction sites to adsorb an extraordinary number of gas molecules.

From the results, N₂O shows the highest gas response due to its strong interaction with the SnO₂ surface, which facilitates effective charge transfer, compared to CO₂ and CH₄. Conversely, greenhouse gas CH₄ displayed weaker responses, with CO₂ showing the least sensitivity. Moreover, the linear increase in sensitivity with concentration suggests that the SnO₂ nanostructures can be reliably used in quantitative gas-detection applications.

2. Analysis of SOV for greenhouse gases

Figs. 6 and 7 demonstrate the output voltages from the MQ4, MQ135, and MQ7 sensors as a function of gas concentration. DAQ recorded these output voltages every 3 seconds. By monitoring the voltage variations, we were able to verify the responses of every gas sensor.

Although it accurately detected CH₄ gas, the MQ4 gas sensor also reacted to CO₂, leading to a similar problem to that of the MQ7 sensor. As a result, these sensors exhibit poor selectivity and hence have problems with gas signal combination. Therefore, it is not feasible to accurately measure the quantities of CH₄, CO₂, or MQ135 gases with a single sensor.

Fig. 6. SOVs for greenhouse gas CO₂ detection.

Fig. 7. SOVs for greenhouse gas CH₄ detection.

5. Analysis of ablation for greenhouse gas categorization accuracy

Involvement of the sensing material: The sensor material's inherent gas sensitivity and initial cross-sensitivity properties are mostly determined by it. Material-level selectivity alone is not sufficient to reliably distinguish gases under changing temperature and humidity conditions, even when the material shows partial selectivity for CH₄, N₂O, and CO₂ due to variations in adsorption and reaction kinetics.

Involvement of the RL algorithm: By actively optimizing operating parameters, including heater temperature, duty cycle, and sampling period in response to environmental circumstances, the RL agent increases selectivity. The RL method reduces misclassification due to cross-sensitivity and environmental degradation by improving class separability in the feature space by learning gas-specific response patterns across various operational conditions.

Involvement of the multiple sensors: Instead of depending on a single sensor response, pattern-based categorization is made possible by the multi-sensor array's complementary response characteristics across sensors. When paired with ML based categorization, this spatial and functional diversity greatly increases robustness and selectivity.

Analysis of Ablation: We conducted an ablation investigation using the following configurations to distinguish these effects: single sensor without RL, multi-sensor without RL, single sensor with RL, and multi-sensor with RL (Proposed system). Fig. 8 shows an accuracy analysis for greenhouse gas categorization using sensors with and without the RL algorithm.

Fig. 8. Accuracy analysis for greenhouse gases categorization using sensors with and without the RL algorithm

The results show that baseline selectivity is influenced by both single and multiple sensors, but that applying RL-based adaptive control to the multiple-sensor array yields the greatest improvement. This demonstrates that the synergistic relationship between adaptive RL control and sensor array diversity drives increased selectivity.

By offering a variety of response patterns, the multiple sensors increase selectivity and produce a discernible increase in accuracy. Class separability is greatly improved by RL-based adaptive control alone by optimizing operating conditions. Applying RL control to the multi-sensor array yields the best results, demonstrating that algorithmic adaptability and sensor diversity work in concert to increase selectivity rather than material attributes alone.

6. Analysis of comparison of baseline experiments using SVM, KNN, and ANN classifiers accuracy

Assuming stable sensing conditions and a stationary data distribution, traditional supervised models such as SVM, KNN, ANN, and CNN are efficient static classifiers. However, due to temperature fluctuations, humidity drift, sensor ageing, and power limitations, sensor responses in gas-sensing systems are highly non-stationary. In these situations, supervised models cannot directly influence the data-collection process and can only learn a mapping from sensor attributes to gas labels.

RL is used as an adaptive control mechanism that operates before classification rather than merely as a substitute classifier. RL makes it possible for:

Optimization of closed-loop sensing: In contrast to supervised models, the RL agent dynamically modifies the sampling interval, duty cycle, and heater temperature depending on the observed sensor condition.

Selectivity and power efficiency: The RL reward function explicitly takes into account both categorization performance and power depletion, allowing accuracy power trade-offs, in contrast to SVM/KNN/ANN/CNN, which optimize categorization accuracy alone.

Adjustment to change in the environment: Without retraining the classifier, RL improves resilience by learning gas-specific optimum operation rules across a range of temperature and humidity circumstances. Fig. 9 illustrates the accuracy comparison for greenhouse gases categorization using KNN, SVM, ANN, CNN and RL algorithms.

Fig. 9. Accuracy comparison for greenhouse gases categorization using KNN, SVM, ANN, CNN and RL

Because it allows for adaptive, closed-loop optimization of sensor operating conditions, something that conventional supervised classifiers are unable to accomplish, the RL is crucial to this study. RL enhances the quality of the measurements themselves, resulting in greater selectivity, resilience, and energy efficiency under equal assessment settings, whereas traditional models improve categorization given fixed measurements.

V. CONCLUSIONS

The integration of SnO₂ nanostructure-based gas sensors with the RL algorithm provides a capable way for accurate, adaptive, and categorization of greenhouse gases, for instance, CO₂, CH₄, and N₂O. The distinctive nanostructure of SnO₂ enhances surface reactivity and gas adsorption, thus enhancing sensitivity and selectivity, while RL enables the proposed mechanism to learn optimal categorization schemes via collaboration with the environment. The greenhouse gas sensing valuation revealed that the SnO₂ nanostructures exhibit remarkable sensitivity and selectivity towards CO₂, CH₄, and N₂O at an optimal operational temperature of 250 °C. By perceptively controlling sensor functioning parameters, the RL agent enhanced accuracy and reached greenhouse gas categorization faster. The experimental results prove that the RL algorithm categorizes with 0.98 accuracy ratio and below 0.1 RFN and RFP. In the future, the proposed system will increase the use of additional gases and mixtures to address cross-sensitivity via federated learning. In addition, the RL algorithm is implemented on hardware platforms for real-world deployment.