Modern Trends in the Application of Electronic Nose Systems: A Review

Abstract

Electronic nose (e-nose) systems have emerged as transformative tools for odor and gas analysis, leveraging advances in nanomaterials, sensor arrays, and machine learning (ML) to mimic biological olfaction. This review synthesizes recent developments in e-nose technology, focusing on innovations in sensor design (e.g., graphene-based nanomaterials, MEMS, and optical sensors), drift compensation techniques, and AI-driven data processing. We highlight key applications across healthcare (e.g., non-invasive disease diagnostics via breath analysis), food quality monitoring (e.g., spoilage detection and authenticity verification), and environmental management (e.g., pollution tracking and wastewater treatment). Despite progress, challenges such as sensor selectivity, long-term stability, and standardization persist. The paper underscores the potential of e-noses to replace conventional analytical methods, offering portability, real-time operation, and cost-effectiveness. Future directions include scalable fabrication, robust ML models, and IoT integration to expand their practical adoption.

1. Introduction

In the last two decades, electronic nose (e-nose) systems have become increasingly important in many fields of science, industry, and daily life. These systems are multisensory devices developed to work similar to human olfaction by using an array of gas sensors in combination with data processing algorithms and recognition and classification models. They are able to record and analyze complex mixtures of gases and volatile organic compounds (VOCs) by extracting a unique “fingerprint” of the smell of a given sample.

The development of electronic noses is closely related to advances in materials science, microelectronics, and artificial intelligence. Thanks to these achievements, modern electronic noses offer high sensitivity, relatively good selectivity, and real-time operation capabilities, making them extremely useful for a wide range of applications. For example, in the food industry, they could be used to control the quality and freshness of products, in medical diagnostics to detect a disease by examining breath or other biological samples, monitoring pollution of an environment, security for detecting explosive substance or drugs, and in the agricultural sector for analyzing soil and plant conditions.

Electronic noses are expected to replace or complement traditional analytical methods, which are often expensive, slow, and require complex sample preparation. With their automation, portability, and on-site capabilities, electronic noses can enable rapid and inexpensive analyses, which is essential for many critical applications.

The aim of this review is to systematize and analyze the existing applications of electronic nose systems in recent years, exploring their core principles of operation and the technologies used. While the present review does not claim to be exhaustive, it provides specific evidence of the broad application of electronic noses and their increasingly sophisticated design.

2. Trends in Structure and Functioning of E-Noses

Gas sensors and electronic noses (e-noses) have undergone transformative advancements in recent years, driven by synergistic innovations in nanomaterials, machine learning (ML), and drift compensation methodologies. Traditional metal oxide sensors (MOS), while widely used, remain constrained by limitations such as long-term drift, poor selectivity, and high operational energy demands [1]. To address these challenges, low-dimensional nanomaterials have emerged as a revolutionary alternative, offering room-temperature functionality, superior selectivity, and energy-efficient operation. For example, graphene-based sensors functionalized with self-assembled peptide films emulate biological olfactory systems, achieving selective detection of plant-derived volatile organic compounds (VOC) such as limonene and menthol through distinct electrical response patterns [2]. Similarly, nanostructured metal oxides, with their rapid response and tunable surface properties, are increasingly integrated into electronic nose architectures for applications that include environmental monitoring, industrial safety, and medical diagnostics [3].

A persistent challenge in gas sensing is sensor drift, which compromises long-term reliability. Recent advancements leverage machine learning (ML) algorithms for real-time drift compensation. For example, hybrid methodologies combining random forest learning with metaheuristic optimization have demonstrated significant improvements in accuracy while minimizing computational overhead [4]. At the decision level, the Gaussian Deep Belief Classification Network integrates unsupervised pre-training and supervised fine-tuning to autonomously correct drift in e-nose systems, outperforming conventional calibration techniques [5]. Beyond drift mitigation, machine learning plays a pivotal role in the decoding of complex sensor data. Supervised learning, unsupervised clustering, and neural network architectures enable precise gas classification and concentration prediction, even in dynamic environments [6]. Transfer learning further enhances adaptability, reducing calibration times for MOS by up to 93% by transferring knowledge across similar sensing environments [7]. Recent reviews underscore the critical role of machine learning in advancing feature extraction, predictive modeling, and drift compensation, with applications extending to robotics, food quality control [8,9], and precision medicine [10]. These methodologies not only improve prediction accuracy, but also ensure long-term stability in diverse operational contexts.

Achieving selectivity, particularly in heterogeneous gas mixtures, remains a difficult obstacle. MEMS-based sensors tailored for specific analytes (eg, CO, NOX, NH₃, HCHO) coupled with principal component analysis (PCA) have proven effective in differentiating binary gas mixtures, highlighting their potential for real-time vehicular emission monitoring [11]. Temperature-modulated MOS, when integrated with convolutional neural networks (CNN), enhance the extraction of features from transient response profiles, enabling the robust discrimination of structurally similar BTX isomers (benzene, toluene, xylenes) [12]. Hybrid systems, such as optical e-noses employing liquid crystal/ionic liquid droplets, achieve 95% classification accuracy for hydrocarbons, although challenges in reusability for some optical e-noses persist [13]. Innovations in optical sensing architectures, including infrared-based modules with Fabry–Perot interferometers (FPI) and pyroelectric detectors, generate distinct voltage patterns for individual gases and mixtures, with PCA effectively clustering responses for integration into solid-state e-nose platforms [14]. Additionally, graphene varactor arrays (comprising 108 sensors) paired with ensemble ML models such as Bagged Random Forest achieve 98% accuracy in VOC identification, even under noisy signal conditions [15].

Polyaniline-based sensors doped with F4TCNQ exhibit enhanced reproducibility in detecting biomarkers like acetone and ammonia, though humidity interference necessitates meticulous calibration [16]. Miniaturized e-nose systems that incorporate optical biomaterials and support vector machines (SVM) achieve a precision of 94.6% in distinguishing 11 disease-related VOCs, assisted by recursive feature selection algorithms [17]. For environmental monitoring, nanowire sensor chips (SnO₂, WO₃, Ge) enable the selective detection of CO, NO₂, and humidity, PCA facilitating the clear clustering of analyte-specific responses [18]. Wearable e-noses, manufactured through the embroidery of polymer/single-walled carbon nanotube (SWCNT) nanocomposites onto textiles, demonstrate potential for personalized healthcare by discriminating individual body odors (e.g., urine, breath) through simplified pattern recognition [19].

Despite progress, trade-offs between sensitivity, stability, and cost persist. Electrochemical sensors maintain superiority over MOS variants in formaldehyde detection, even after refinements of ML-enhanced accuracy [20]. Field studies reveal persistent long-term drift in e-nose systems, though calibration transfer methodologies sustain NO₂ detection accuracy over six-month deployments [21]. Future research must prioritize scalable nanofabrication techniques, such as anodic aluminum oxide (AAO) templating, to enable cost-effective mass production of sensor arrays [22]. Although boosting algorithms (eg, XGBoost) surpass deep learning in computational efficiency and robustness for gas recognition tasks [23], ensuring signal independence across sensor arrays remains critical to mitigate cross-sensitivity artifacts [24]. Emerging architectures, such as gold nanoparticle-enhanced e-noses with near-perfect VOC quantification [25] and AI-driven systems achieving 100% accuracy in coffee origin classification [26], underscore the transformative potential of smart sensing technologies. However, broader adoption is based on resolving standardization gaps, enhancing drift resilience, and optimizing manufacturing economies of scale. One study [27] presents the developed graphene-based e-nose functionalized with aryl groups for selective detection nitrogen dioxide (NO₂), demonstrating high sensitivity (1–10 ppm) and over 95% classification accuracy using PCA and linear discriminant analysis. Another system [28] uses tin oxide semiconductor sensors to detect VOCs such as benzene, acetone, and ethanol, forming unique smell prints for applications in pharmaceuticals, defense, and security, with strong selectivity and reliability. Comprehensive reviews [29,30] highlight the general principles of e-nose systems, which mimic human olfaction using sensor arrays and pattern recognition. They emphasis the versatility of e-noses in food quality control, medical diagnostics, and environmental monitoring. Key advancements include the use of diverse sensing materials, including metal oxides, carbon-based compounds, and hybrid structures, and innovations in bilayer sensor designs and high-throughput screening. The convergence of advanced nanomaterials, AI-driven signal processing, and innovative sensor architectures is redefining the capabilities of gas sensing systems. However, persistent challenges in sensor longevity, cross-sensitivity, and cost-effective production demand interdisciplinary collaboration to unlock their full impact. Future efforts must harmonize material innovation with algorithmic robustness to realize a scalable, reliable, and affordable smart sensing solution.

When analyzing the topic of the e-nose, three areas should be distinguished. As shown in Figure 1, the most basic are gas sensors that allow information to be collected. They are used by various Machine Learning Models. However, how they are trained depends on the intended use of these systems.

This article focuses on the most common areas of e-noses applications. However, this requires taking into account Machine Learning Models. In order to understand the principle of their operation, they require knowledge of gas detection sensors used by e-noses. Therefore,

Table 1. Gas sensors type overview.

Sensor Type	Principle of Operation	Purpose (Application and Gases)	Size of a Sensor	Advantages	Disadvantages
Metal Oxide (MOS/MOX) [31,32,33,34]	Change in electrical resistance in a metal oxide layer upon gas adsorption	Air quality, food, medical VOCs, CO, NOX, NH3, H2, alcohols	Small to very small	Low cost, fast response, miniaturization	Sensor drift, low selectivity, sensitive to humidity
Conductive Polymer [35]	Change in conductivity/capacitance in polymer upon VOC exposure	Food quality, environmental VOCs, solvents	Small/medium	Flexible design, many target VOCs, low cost	Aging, drift, limited stability
Electrochemical [20]	Current generated by redox reaction at electrode in contact with analyte gas	Industrial safety, environmental, toxic gases (CO, SO2, NO2)	Medium	Good selectivity, linearity, low power	Finite lifespan, more costly, cross-interferences
QCM (Quartz Crystal) [36]	Change in resonance frequency due to mass of adsorbed gas molecules	Trace VOCs, environmental, medical, research	Medium	Very high sensitivity, can detect ppb	Sensitive to temp/humidity, complex, more costly
Optical (e.g., NDIR) [37]	Absorption of specific IR wavelengths by target gas in optical cell	CO2, CH4, refrigerants, air monitoring, safety	Large/medium	Very selective, stable, requires little maintenance	More expensive, larger size, not for all gases
Acoustic (SAW/BAW) [38,39,40,41,42,43,44,45,46,47,48,49,50,51,52]	Shift in surface/bulk acoustic wave caused by gas adsorption on sensor surface	VOCs, industrial processes, research	Small/medium	High sensitivity, fast response	Sensitive to environmental changes, complex packaging
Biosensors/Biomimetic [53,54]	Interaction of volatile molecules with biological (or biomimetic) receptors	Medical diagnostics, specific biomarker VOCs	Small/medium	Ultra-high selectivity, unique specificity	Fragility, complexity, stability, cost

presents a synthetic summary of different types of sensors and their application areas, advantages, and limitations. In addition,

Table 2. Machine Learning techniques overview.

Technique	Category	Strengths	Use Cases in Manuscript
PCA [11]	Unsupervised	Dimensionality reduction, clustering	Breath analysis, food spoilage
SVM [17]	Supervised	Effective on small datasets	Diabetes, pneumonia, asthma
Random Forest [4]	Ensemble	High robustness to noise	Food fraud, VOC classification
CNN/1D-CNN [55]	Deep Learning	Temporal/spatial signal feature extraction	Asthma detection, meat classification
Autoencoders [10]	Unsupervised	Denoising, anomaly detection	Air quality drift correction
XGBoost [23,56]	Boosting	High accuracy, interpretable	Ammonia, acetone detection
Transfer Learning [7]	Cross-domain	Reduces calibration effort	Gas classification in new contexts
LDA (Linear Discriminant Analysis) [57]	Supervised	High interpretability of results. Effective for small and medium-sized datasets. Low computational requirements. Resistant to overfitting with proper feature selection. Easy implementation	Classification of garlic aroma profiles based on e-nose data to distinguish processing methods; enhanced discrimination of different garlic treatments

contains a summary of the Machine Learning techniques used, the strengths of each method, and the use cases presented in the article. Sensing materials used in gas sensors are presented in

Table 3. Sensing materials overview.

Sensing Material	Structure Type	Fabrication Technique	VOCs Detected
Graphene (e.g., aryl-functionalized) [27]	Single-layer 2D nanomaterial	Diazonium chemistry, CVD, laser-induced	NO2, benzene, acetone, ethanol
SWCNT/polymer composite [19,35,58,59]	Nanocomposite, 1D+polymer	Solution casting, drop casting	Acetone, ammonia, H2S, ethanol, body/breath VOCs
Metal oxide nanowires (e.g., SnO2, WO3, Ge) [18,60]	1D nanowire array	Glancing angle deposition, thermal oxidation	CO, NO2, CH4, NH3, acetone, CH3OCH3
MXene/WO3 hybrid [61]	Layered 2D–3D nanohybrid	Solution-based assembly, drop casting	Toxic gases (NO2, NH3, CO, CH4, acetone)
QCM—coated with ethyl cellulose, PMMA, Apiezon [62]	Film-coated quartz crystal	Spin-coating, drop-casting on QCM	Acetone, mixed VOCs (diabetes biomarkers)
ZnO–peptide hybrid [2,63,64]	Functionalized nanostructure	Self-assembly, peptide functionalization	Limonene, menthol, plant VOCs
MOS (e.g., TGS, MQ-series, BME688) [1,64,65]	Granular/pellet/film arrays	Thick-film, screen-printing, MEMS integration	H2, CO, CH4, NOx, NH3, food VOCs
Optical (Fabry–Perot, pyroelectric) [14]	Thin film, IR module	E-beam evaporation, FPI integration	Hydrocarbons, BTX (benzene, toluene, xylenes)
Polyaniline/F4TCNQ-doped [16]	Conducting polymer composite	Chemical doping, drop casting	Acetone, ammonia, formaldehyde
Peptide-based (for QCM, FET) [66,67]	Peptide monolayer or film	Self-assembly, drop-casting	Carrot VOCs, other food markers

.

The technologies, materials, and solutions presented above have been implemented in commercially available devices. They range in complexity from simple detectors of selected gases to versatile devices such as electronic noses capable of recognizing multiple substances and matching them to specific odors.

At this point, it is worth mentioning general trends among various sensors that could serve as future inspirations for e-noses. The first is ‘flexible sensors,’ which are characterized by the ability to be crushed and deformed, allowing them to adapt to various surfaces, such as skin, without losing their properties [68]. Another direction is ‘self-powered sensors.’ These utilize voltage induction phenomena (as in RFID) or piezoelectricity [69]. This eliminates the need for a battery that would require replacement. Another is combining sensors with AI, which has already been discussed in this section. Section 3.1 will describe the use of sensors in IoT devices. However, this trend is directly related to the miniaturization and portability of these devices [70].

3. Areas of Application

3.1. E-Noses in Healthcare

Electronic nose technology has already revolutionized non-invasive diagnostic methodologies in modern healthcare. By leveraging the detection of volatile organic compounds (VOCs) in breath, urine, and blood, e-nose systems present a paradigm shift from conventional invasive, costly, and time-consuming diagnostic techniques. Recent innovations in sensor materials, machine learning (ML), and artificial intelligence (AI) have substantially augmented the precision, portability, and clinical applicability of these devices, enabling their integration into diverse medical domains.

E-nose systems have shown exceptional efficacy in oncology, particularly in early cancer detection. A compact e-nose incorporating eight polymer-functionalized single-walled carbon nanotube (SWCNT) sensors achieved high sensitivity in distinguishing patients with hepatocellular carcinoma (HCC) from healthy individuals through breath analysis, using principal component analysis (PCA) for dimensionality reduction [35,71]. Further advancements in AI-driven systems have elevated diagnostic accuracy; for instance, a five-sensor e-nose employing linear discriminant analysis (LDA) attained 93.14% accuracy, 88.63% sensitivity, and 95.62% specificity in lung cancer detection via breath VOC profiling [72]. Diversification of sensors has proven to be critical, as evidenced by a 10-sensor e-nose system that achieved an area under the curve (AUC) of 0.87 for the diagnosis of lung cancer using a Random Forest classifier [73]. These studies underscore the synergistic potential of multi-sensor arrays and advanced ML algorithms in oncology.

E-nose technology has also emerged as a tool for diabetes monitoring, particularly in identifying metabolic biomarkers. A low-cost, portable e-nose with four polymer/SWCNT nanocomposite sensors demonstrated near-perfect accuracy (99.5%) in classifying urine samples from type 2 diabetes (T2D) patients, with pronounced responses to acetone, ammonia, and ethyl methyl ketone [58]. Researchers have also developed a system [74] which use six chemiresistive sensors to analyze urine samples, effectively distinguishing between diabetic and healthy individuals through VOC profiles using PCA and cluster analysis. Another system [60] employed a metal oxide sensor array to analyze breath biomarkers, achieving 90.4% accuracy in blood glucose monitoring with machine learning models. Innovations in sensor design include a quartz crystal microbalance (QCM)-based array coated with ethyl cellulose (EC), polymethylmethacrylate (PMMA), and Apiezon films, which achieved 100% classification accuracy for acetone—a key diabetes biomarker—across concentrations spanning 327–4908 ppm [36]. Complementing these findings, a smart toilet-integrated e-nose system utilizing eight metal oxide sensors (e.g., TGS2603, MQ8) successfully discriminated diabetic urine VOCs via PCA, highlighting the feasibility of continuous at-home monitoring [75]. Metabolomic profiling via ultra-fast gas chromatography e-nose (FGC eNose) further identified T2D-specific urinary VOCs, including 2-propanol and butane-2,3-dione, achieving 84.6% classification accuracy in a clinical cohort [76].

In cardiovascular care, a 19-sensor e-nose system distinguished myocardial infarction patients from healthy controls with 97.19% accuracy using a support vector machine (SVM) classifier, while attaining 81.48% accuracy for stable coronary artery disease [77]. For respiratory conditions, e-noses have shown remarkable utility in chronic obstructive pulmonary disease (COPD) diagnosis, with one system achieving 88.57% accuracy via PCA and cluster analysis [56], and an FGC eNose identifying 17 COPD-specific VOCs with 96% sensitivity and 91% specificity [78]. Expanding to asthma, a genetic algorithm-optimized e-nose reduced sensor redundancy by selecting five optimal sensors, paired with a 1D convolutional neural network (1D-CNN) model, achieving 96.6% accuracy in discriminating asthma patients via exhaled breath analysis [55].

E-nose systems have been validated for rapid infection diagnostics. A 28-metal oxide semiconductor sensor array (MOS) combined with SVM and artificial neural networks (ANN) differentiated ventilator-associated pneumonia (VAP) patients with 92.08% and 85.47% precision, respectively [31]. Hybrid optical sensors using thin films of zinc porphyrin and manganese porphyrin successfully discriminated between bacterial species (S. aureus, E. coli, P. aeruginosa) during different growth phases [79].

Recent breakthroughs in e-nose architecture include modular, wireless platforms utilizing MXene-based porous structures and laser-induced graphene electrodes, which achieved 91.7% accuracy in disease diagnosis via urine VOC analysis [80]. IoT-enabled wearable systems, such as a MOX sensor-based e-nose (like on Figure 2) with neural networks, demonstrated 100% accuracy in real-time synthetic urine detection, paving the way for remote patient monitoring [81]. Additionally, hybrid systems integrating SVM and least-square regression models have been deployed for refinery gas classification and hemodialysis monitoring, illustrating cross-industry adaptability [82].

Despite these advancements, persistent challenges include sensor drift, environmental interference (e.g., humidity, temperature), and the need for large-scale, multi-center clinical validation [83]. Future research must prioritize the development of ultra-selective sensor materials, such as molecularly imprinted polymers (MIPs), and the integration of multi-modal data (e.g., VOCs with genomics or proteomics) to enhance diagnostic specificity. Additionally, optimizing ML algorithms for real-time applications—particularly in resource-limited settings—remains critical, as highlighted by studies on genetic algorithm-driven sensor optimization [55] and portable diabetes monitoring systems [36,75].

3.2. Food Detection and Classification

E-nose systems demonstrate exceptional performance in spoilage detection. Gas sensor-based e-noses paired with ANNs achieve 99.58% accuracy in classifying apple rot caused by fungal pathogens [84], while semiconductor sensors streamline tomato quality assessment [85]. Recent advancements include SnO₂ nanopetals sensors for detecting formalin-adulterated fish [86] and PEN3 e-nose systems distinguishing moldy apples (Penicillium expansum, Aspergillus niger) with 96.3% accuracy [87]. For exotic fruits (durian, jackfruit, mango), chromatographic column-based e-noses achieve 82% classification accuracy [88]. ZnO-peptide QCM sensors predict carrot shelf life under varying storage temperatures [89], and portable systems with DSP processors classify apple odors using back-propagation neural networks [90]. Multi-sensor fusion systems can classify seven fruits and eight vegetables with 96.7% accuracy [91], while nanocomposites enhance formaldehyde detection in formalin-laced foods [92].

Low-cost e-noses can successfully distinguish fresh from spoiled meat (>95% accuracy) [93,94], and CNN-based systems with MEMS sensors achieve 98.4% accuracy by combining transient/steady-state odor data [95]. Hydrogen sulfide tracking predicts poultry freshness over 48 h [96,97], while MQ135/MQ4 sensors combined with AI image processing overcome challenges from artificial meat coloring, achieving >90% accuracy [98].

Electronic noses with Bosch BME688 sensors classify post-opening milk spoilage [99]. Semiconductor nanowire sensors authenticate Parmigiano Reggiano cheese by VOC profiles and aging (100% accuracy) [65], supported by hierarchical ANN models (88.24–100% accuracy for rind percentage and aging) [53]. A novel portable sensor device using six MOS nanowire sensors [54] successfully distinguished quality and ripening stages of Parmigiano Reggiano cheese through VOC analysis and PCA. Similarly, a portable electronic nose with an MQ137 sensor [100] monitored ammonia levels to assess Asian green mussel freshness, achieving 95% classification accuracy over a 60 h spoilage period. Comprehensive reviews [101,102,103] further explore the application of MOS and MOS-resistive (MOSR) sensors in evaluating freshness and spoilage in fish and meat. These studies examine chemical changes during decomposition, signal processing techniques, and pattern recognition methods.

E-noses combat fraud through 4-sensor systems that differentiate Malaysian vs. Vietnamese black pepper (100% accuracy) [104] and detect adulterants (e.g., pumpkin, starch) in tomato sauce via metal-oxide sensors [105]. For potatoes, SVM classifiers identify cultivation origin (organic/chemical/manure) with 91.7–96.7% accuracy [106]. MOS sensor arrays authenticate Sarawak black pepper (84.8% accuracy) [107], while MOGS arrays classify chocolate type and expiration status (81.3% accuracy) [108].

In beverages, e-noses classify black tea aroma (88.9% accuracy) [109] and distinguish artificial strawberry flavors using polyaniline/carbon nanotube sensors [110]. For canned tuna, f-SWCNT sensors monitor spoilage via ammonia emissions (PCA visualization) [59]. A wearable e-nose with four sensors classifies liquid food intake using ANN, Random Forest, and K-NN [111], while systems with seven MQ sensors detect spoilage in tomato-based Filipino dishes (96.15% accuracy) [112]. Coffee roasting stages can also be monitored via gas emissions [113].

Advances in food detection and classification include bimetallic oxide sensors with CNN-LSTM models for gas quantification (RMSE 2.3) [114], 20-sensor e-noses optimized to six sensors for 90% biscuit classification [115], and portable systems for farm product monitoring [116]. Deep learning models classify tomato ripeness (86% accuracy) [117]. Researchers also developed an electronic nose with temperature-modulated sensors and machine learning (SVM/RF algorithms) to detect salinity stress in Khasi Mandarin Orange plants via VOC analysis, achieving 98.3% accuracy, validated by chlorophyll and fluorescence measurements [118].

Figure 2. Portable e-nose for detection of Carbon Monoxide (CO), Nitrogen Dioxide (NO₂), Ammonia (NH₃) based on MICS-6814 gas sensor.

Figure 2. Portable e-nose for detection of Carbon Monoxide (CO), Nitrogen Dioxide (NO₂), Ammonia (NH₃) based on MICS-6814 gas sensor.

According to many studies, i.e., [119,120], in recent years, a special role has been given to the use of sensors and data processing algorithms to check product quality and ensure safety, both in industrial and food contexts.

3.3. Environmental Monitoring

One of the studies in this field proposes a new method for gas recognition and concentration prediction using a Time–Frequency Attention Convolutional Neural Network (TFA-CNN), which allows for better integration of temporal and frequency information derived from sensor signals [121]. The inclusion of a specialized attention block and data augmentation strategy significantly improves the model’s robustness to sensor drift and redundant information. Experimental results with MOS on five different gases demonstrate 100% classification accuracy and a determination coefficient (R²) of 0.99 in the regression task, confirming the system’s high performance under real-world conditions.

A review [122] discusses recent advances in air quality monitoring tools—such as low-cost sensors, gas analyzers, passive samplers, and remote sensing methods (e.g., satellites, LiDAR)—for detecting pollution from natural and human-made sources. It emphasizes the need for improved calibration, standardization, and data integration to ensure accurate, real-time monitoring aligned with public health policies. Exemplification can be provided by studies [123,124], among others. One review [125] focuses on the pulp and paper industry, where harmful gases are emitted during production. It evaluates the use of electronic noses, particularly metal oxide semiconductor (MOS) sensors combined with artificial neural networks (ANNs) and highlights the ZigBee protocol for wireless sensor networks. In practice, wireless means of communication, including satellites, can be used to transmit data from remote and hard-to-reach locations [126,127,128,129,130,131,132,133,134,135,136,137]. These techniques can be useful for post-crisis management in the broadest sense and for providing security in social terms [138,139,140,141,142]. Modern vision techniques can be used to detect flames [143,144,145,146,147,148,149,150], while acoustic waves can be used to extinguish them [38,39,40,41,42,43,44,45,46,47,48,49,50,51,52], which opens a new stage in the development of fire protection techniques and supports broader security.

In [151] an electronic nose (i.e., e-nose) using a nanostructured chemical sensor array (hierarchical/doped design) is applied to detect seven gases (H₂, C₂H₂, CH₄, CH₃OCH₃, CO, NO₂, NH₃) with high sensitivity and rapid response. Artificial neural networks (ANNs) optimize sensor linearity, mitigate temperature effects, and enable fast qualitative/quantitative gas discrimination, demonstrating its potential for environmental monitoring applications.

Another important contribution to gas sensing is discussed in [152], where traditional limitations of graphene sensors under atmospheric conditions are overcome by passivating the graphene channel with activated carbon and applying machine learning (XGBoost, KNN, and Naïve Bayes). Using SHapley Additive exPlanations (SHAP) analysis to identify the most significant features, it is shown that the effects of van der Waals doping and gas scattering change over time. As a result, the models achieve up to 100% accuracy in detecting ammonia and acetone, even at very low concentrations (ppb level), opening up opportunities for practical application outside laboratory conditions.

Electronic olfaction systems are also applied in environmental process management. In [153], an e-nose is presented that combines a hardware sensor system and machine learning to classify the stages of wastewater treatment. The use of techniques such as t-SNE for dimensionality reduction and k-median for clustering enables a successful differentiation of the chemical signatures of each stage. A Random Forest model demonstrates perfect classification without signs of overfitting, confirming the system’s potential for automated wastewater treatment plant management.

Equally important are efforts to improve the selectivity and accuracy of semiconductor metal oxide (SMO) sensors. In [154], a unified sensor array of metal oxides (SnO₂, In₂O₃, WO₃, CuO), developed via glancing angle deposition, combined with deep learning (CNN), achieves 98% accuracy in classifying five different gases (CO, NH₃, NO₂, CH₄, and acetone). The system shows a very short response time (from 1 to 19 s), which is critical for real-time monitoring of the environment.

The focus on biological applications is also increasing. Studies in [155,156] use low-cost e-noses with TGS sensors for identifying fungal pathogens. In [155], it is demonstrated that modulation of the heater voltage provides more informative features for training models, especially in the initial stages of modulation. In [156], the rectangular modulation approach of heater voltage shows advantages in continuous seed monitoring, achieving a classification accuracy of 97% through a combination of PCA and Random Forest.

In the field of volatile organic compound (VOC) recognition, the virtual e-nose presented in [157] offers an innovative approach using quadratic support vector machine (SVM) and partial least squares (PLS) regression, achieving 91% accuracy and R² up to 0.99. This provides an alternative for addressing the longstanding issue of sensor selectivity.

An interesting example of e-nose application in agroecology is presented in [158], where a system with 10 identical sensors operates at different temperatures to predict soil organic matter (SOM) content. The support vector regression (SVR) model achieves the best results (R² = 0.895), demonstrating the potential of e-noses in soil monitoring. Additionally, Ref. [159] presents a portable and energy-efficient e-nose with 16 miniature MOX sensors (BME680), which maintains accuracy above 70% even two weeks after calibration, making it suitable for mobile and battery-powered applications [160].

Innovations with MXene-based sensors, presented in [61], show enhanced sensitivity to NO₂ gas through the use of a memristive spiking neural network, achieving 95.83% accuracy. These sensors demonstrate excellent stability and reliability for long-term use.

The study in [161] introduces a method to detect acetone by analyzing fluctuations in the resistance of gas sensors, with entropy identified as the most informative feature. The model demonstrates good robustness even in mixed gas/noise environments. Industrial applications of e-noses are observed in [162], where the technology is used to analyze odors from damaged oils and cables, allowing for early fault detection. In [163], an automotive safety system is examined, which automatically detects toxic gases inside the vehicle cabin and takes protective actions.

Finally, Ref. [164] demonstrates the successful application of e-noses for indoor air quality monitoring, where the neural networks used can classify and predict the concentrations of hydrogen, methane, and carbon monoxide with high accuracy, even in the presence of interfering gases.

The decreasing cost of sensors detecting individual gases combined with advanced information processing makes these solutions more and more accessible. For example, a low-cost e-nose sensor head can be used in various types of environmental monitoring applications.

3.4. Key Application Requirements for E-Nose Systems

The requirements for electronic nose (e-nose) systems depend on their specific area of application.

Table 4. Summary of domain-specific requirements.

Area of Application	Key Requirements	Challenges
Healthcare and Medical [36,55,56,58,60,73,77]	High sensitivity and selectivity; non-invasiveness; robust validation; reproducibility	High interpretability of results; clinical relevance; data privacy; resistance to environmental interference
Food Quality and Safety [54,84,85,86,89,90,93]	Detection of freshness/spoilage; selectivity for complex mixtures; rapid analysis; cost-effectiveness	Robustness to matrix effects; traceability; detection of adulteration; sensors stability
Environmental Monitoring [118,121,122,125,127,134]	Rapid response/real-time detection; long-term stability; remote/autonomous operation; wide analyte range	Timely prediction; resistance to sensor drift; operation in harsh/outdoor conditions
Industrial Safety and Security [148,149,150]	Fast detection of hazardous/toxic gases; low false alarm rate; integration with alarm systems	Reliability over time; operation under variable conditions; network capability

presents a concise summary of the domain-specific requirements and special needs for sensors and data analysis methods in the medical, food, and environmental fields.

4. Statistical Information

The overview of current trends in the use of electronic noses is based on a wide range of publications. These publications can be broadly classified according to their type, the methods and approaches used in the operation of the electronic noses described in them, the country of origin of the publication and the field of application to which the respective electronic nose relates.

This review of publications began with the collection of information on articles from the Scopus database. In the first step, the following criteria were applied:

• Abstract of an article contains the phrase “Electronic Nose Detector” or “Electronic Gas Detector”;
• Years of publication from 2015 to 2024;
• Language of publication is English;
• Limited to keywords: “Principal Component Analysis”, “Feature Extraction”, “Support Vector Machines”, “Machine Learning”, “Learning Systems”, “Neural Networks”, “Pattern Recognition” or “Classification”.

The limitation to the given keywords resulted from the desire to focus on the most important and current methods and techniques used in data analysis and machine learning [165]. Since they do not specify a single solution, but concern a whole group of solutions, they reflect the main trends of research. This was to eliminate marginal works or works not directly related to the research topic, which allowed for a more reliable analysis of trends and identification of the most important achievements and research problems.

As a result, 313 publications were found, thus articles were limited to those with the following keywords: Diagnosis, Diseases, Biomedical Engineering, Biomarkers, Environmental Monitoring, Spoilage, Forestry, Air Quality, Quality Control, Fruits, Food Quality, Food Safety, Meats, Chemistry, Chemical Gas Sensors. The result is shown in Figure 3. The keyword density is presented in Figure 4.

4.1. A Publications Overview

Statistical information about the reviewed articles is presented in

Table 5. Publications by year in all categories.

Name	2015–2019	2020–2024	All years	Share [%]
Total	33	70	103	100.0
Document Type
Conference Paper	14	17	31	30.1
Journal Article	18	44	62	60.19
Other	1	9	10	9.71
Data Analysis Methods
Machine Learning	6	47	53	51.46
Neural Networks	11	18	29	28.16
Feature Engineering	17	20	37	35.92
Pattern Classification	10	27	37	35.92
Electronic Nose Applications
Food Agriculture	12	23	35	33.98
Medical Biomedicine	10	23	33	32.04
Material Chemistry	9	8	17	16.5
Environmental Monitoring	7	24	31	30.1
Air Monitoring	1	11	12	11.65
Research Methodology
Experiment	31	60	91	88.35
Literature Analysis	5	15	20	19.42
Case Study	2	3	5	4.85
Conceptual	9	24	33	32.04

. The period under study was divided into two decades. Additionally, the summarized information and share of them in total amount were included.

The articles were divided using a few criteria. Thus, there were four groups created. In the groups “Data Analysis Methods”, “Electronic Nose Applications”, and “Research Methodology” an article could belong to more than one category, thus values in column “Share” do not sum up to 100%. In the next subparagraphs, they will be analyzed.

4.2. Types of Publications Reviewed

The publications reviewed can be divided into two groups depending on the year of their publication—for the period 2015–2019 and for the period 2020–2024. Figure 5 presents a graph depending on the places where the publications reviewed were published, separated depending on the year of publication. Articles in conferences are 14 (for the period 2015–2019) and 17 (for the period 2020–2024), respectively. Articles in journals are 18 for the first period and 44 for the second period of publication, respectively. The remaining publications are placed under the column “other”.

Depending on the research methodology, the number of publications is presented graphically in Figure 6. The individual publications are divided into the following main groups: Experiment, Literature Analysis, Case Study, and Conceptual.

4.3. Data Processing Methods Applied in Electronic Noses

The main data processing methods in various implementations of electronic noses include Machine Learning, Neural Networks, Feature Engineering, and Pattern Classification. Mainly, developments using Machine Learning dominate, as they are included in a total of 53 publications for the two considered time periods. The application of artificial neural networks as a modern approach is gaining its application on a pair with such established methods as Feature Engineering and Pattern Classification. Figure 7 presents statistical data on the data analysis methods used in a graphical form.

The data analysis methods used are presented in Figure 8 in a form that shows their relationship with the research methodology used.

Articles were qualified manually, based on their abstracts and contents. The first group, Experiment, contains these papers, where results of experiments were presented, like noise level measurement. If a publication contained already published surveys or reviews, it was included in the second group, which was named as Literature Analysis. The next one, Case Study, collects articles describing solutions for a real problem, where something had to be developed or implemented. The Conceptual group is built by papers, where some models or concepts were developed.

4.4. Authorship of Publications by Country

Figure 9 presents the publications for the two periods under consideration depending on the countries to which their authors belong. The largest number are publications in which the authors are a team of scientists from different countries. In this case, they are summarized under the class “Other”. It is noticeable that publications by authors from countries in the Far East—China, Thailand, India, etc.—dominate. The number of publications from these countries is over half of the total number of publications considered in this review.

4.5. Application Areas of Electronic Noses

In the present work, the literature sources used are categorized according to their main application into the following groups: those related to the trends and structure of electronic noses; those concerning the application of electronic noses in medicine; applications in food quality assessment; and those related to environmental monitoring. Figure 10 presents a graphical representation of the statistics on the application areas of electronic noses discussed in this review.

Although the three groups of applications described above seem to be the most obvious, it can be expected that e-noses will also find application in other areas. For example, it seems that the use of these devices for detecting used equipment (e.g., oxidation of circuits, burnout of electrical circuits) or checking the quality of pig iron in blast furnaces based on fumes has not been studied. These may therefore be new areas of research. A summary is presented in

Table 6. A summary of possible application areas.

Application Area	Example Application	Sensed Compounds
Medicine	Diabetes monitoring from urine/breath	Acetone, ethanol
Food industry	Detection of fish/spoiled meat	Ammonia, H2S, VOCs
Environmental	Pollution monitoring in urban areas	CO, NO2, NH3, VOCs
Security/Material Chemistry	Detection of explosives	TNT, DNT vapors
Air monitoring	Indoor air quality measurement	CO2, VOCs, formaldehyde

.

4.6. Electronic Nose Standards

One of the problems with the use of electronic noses is the lack of standards regarding their use and testing.

Table 7. Examples of standards for electronic noses.

Standard Number	Standard Title	Part (If Applicable)	Purpose
ISO 16000-29 [166]	Indoor air—Determination of volatile organic compounds (VOCs)	Part 29: Measurement of VOCs using electronic noses	Describes the use of e-noses for determining VOC concentrations in indoor air.
ISO 12219-7 [167]	Interior air of road vehicles	Part 7: Odor determination by olfactory measurement with e-nose	Specifies methods for evaluating odors inside vehicles with e-nose technology.
ISO 25140 [168]	Stationary source emissions—Automatic method for the determination of the methane concentration using flame ionisation detection (FID)	—	Provides procedures for the determination of odor concentrations in gases, including e-nose methods.
ISO 16000-29:2014 [166]	Indoor air	Part 29: Test methods for VOC detectors	Specifies performance testing procedures for electronic noses detecting volatile organic compounds.
ISO/IEC 17025 [169]	General requirements for the competence of testing and calibration laboratories	—	General requirements for laboratory quality and competence, relevant to e-nose test/analysis labs.
VDI 3880 [170]	Olfactometry-Static sampling	—	German guideline providing comprehensive procedures for validation and performance evaluation of e-noses.
ISO 13320 [171]	Particle size analysis—Laser diffraction methods	—	Specifies methods for particle size analysis by laser diffraction; not directly related to e-noses, but often cited in laboratory measurement practice.

lists examples of standards used in this area.

The presented standards therefore mainly specify the principles of sensor testing, the way they are used and calibrated. They also specify the sizes of the fraction of suspended particles. However, there is a lack of general design and verification standards for this type of implementation. This may be particularly important in the context of those sensors in which the drift phenomenon occurs and there is a possibility of incorrect detection of the tested odor.

It should be noted that some producers of gas sensors (especially electro-chemical ones) give in their documentation procedures for how the calibration of sensors should be performed. These protocols of calibrations are product specific, and they are not included in the paper.

4.7. Summary of Key Points

When analyzing the articles presented, several key points can be identified. They refer to both the problems encountered and suggestions for solving them. Therefore, they are placed in the following list:

• Selectivity remains a challenge in e-nose systems. MEMS-based sensor arrays combined with PCA or other dimensionality reduction algorithms help improve discrimination by filtering noise and enhancing signal clarity. Using modern techniques as artificial neural networks can also be beneficial to solve problems with cross-sensitivity of gas sensors and to increase the selectivity of e-noses.
• There is always the possibility for over-fitting when ML methods such as artificial neural networks are used. It depends on the structure of the training data as well as the number of artificial neurons used and the structure of the neural network. Often the training of proper neural network is an iterative process during which the researchers have to be able to find eventual over-fitting and to take the necessary measures against it, making additional processing of training data or changing the structure of artificial neural network.
• The GC-MS (Gas Chromatography–Mass Spectrometry) are traditional methods for analyses with very good resolution and accuracy, but they are expensive. The e-noses give low-cost measurements, but they depend on software methods used for data processing to achieve the accuracy of GC-MS.
• Electronic noses distinguish between different diseases based on a set of sensor readings, rather than a single piece of data. These sets of readings are processed using machine learning methods and, thus, using different ML methods, even small fluctuations in the input dataset can detect different diseases.
• Combining e-nose data with spectroscopy (e.g., Raman, NIR) can boost diagnostic confidence. Such multimodal systems improve specificity. For instance, according to some publications, breath analysis combining an e-nose with infrared spectroscopy enhances disease discrimination compared to either method alone.
• Reliability of IoT e-nose systems depends on data communication protocols used, the cryptography of transferred data, algorithms for error compensation, and many other factors. As examples of IoT e-nose systems such devices can be pointed that use the following communication protocols: LoRa/LoRaWAN, Wi-Fi, NB-IoT, Bluetooth, BLE (Bluetooth Low Energy), etc.
• Edge computing allows processing, analysis, and decision-making to occur locally on e-nose systems. The characteristics of edge computing are low latency, reduced bandwidth used for data exchange, and possibility for e-noses to operate without need for connectivity with other devices such as servers and cloud services. Working offline the e-noses with edge computing also provide good privacy.
• Wearable e-noses, as other wearable devices, have to use little energy because the concept of all wearable devices is to be autonomous and to use batteries or energy harvesters as sources of energy. Being battery powered, the wearable e-noses have to use sensors which have low consumption, fast response time, and low weight. There are also different techniques that can optimize the energy consumption of such e-noses; these techniques depend on electronics used (as microcontrollers, voltage regulators, etc.) and also depend on software algorithms for reducing the consumption and increasing the battery life.
• To address the longevity of the gas sensor, interdisciplinary collaboration is essential. Combining materials science and AI can help to discover new and more stable sensing materials and predict the degradation and aging effect using machine learning. Materials science and electrical engineering can work together to create materials and circuits that compensate for sensor drift over time. This collaboration can significantly extend sensor lifespan and reliability in real-world applications.
• The trends in e-noses from the last five years include innovations in integration of artificial intelligence and machine learning in their structure as deep neural networks and ML methods as Random Forests, XGBoost, KNN, and Nave Bayes. Other trends include the increasing use of IoT and edge computing. For the processing of data from e-nose sensors, specialized neuromorphic chips and TPUs (Tensor Processing Units) can be used. Other trends are innovations in the field of sensor materials such as specialized polymers. All mentioned above trends lead to miniaturization of e-nose systems, improving their accuracy and making them a real low-cost alternative of specialized measurement equipment in the fields of healthcare, food industry, and environment monitoring.
• The discovery of new material properties that can be used to build new sensors may also have a significant impact on the further development of e-nose devices. For example, Ref. [172] describes the gas adsorption and sensing properties of a nickel-decorated WS₂-WSe₂ (Ni@WS₂-WSe₂) heterojunction on C₂H₂ and C₂H₄, for dissolved gas analysis and evaluation of oil-filled transformers. This technology enables the construction of highly sensitive, reusable sensors.

5. Conclusions

In recent years, there has been a significant development of electronic nose technology. Using technological improvement in gas detection sensors and machine learning algorithms, significantly higher levels of sensitivity of e-nose systems have been achieved. They are also operating in real time. Therefore, this technology is a tool for analyzing the actual state of affairs in many fields. This review shows that e-nose technology can lead to major changes in many fields, including medical diagnostics, food quality control, and environmental monitoring. They have been shown to be able to not only complement, but even replace traditional analytical methods, which are often more expensive and time-consuming.

E-noses are able to identify quick and accurate volatile organic compounds (VOCs) associated with diseases such as lung cancer, diabetes, and respiratory infections. Thus, this technology can be used in healthcare for non-invasive diagnosis. The ability to analyses breath, urine, and other biological samples with high diagnostic precision allows them to be used for early screening tests. Another area where e-noses have been applied is the food industry. For example, they are able to detect spoilage of a food, verifying its origin, and control production processes, which significantly reduce waste and ensure consumer safety. Due to the ability to detect different types of gases, it is natural to use them for monitoring air pollution, industrial emissions, or optimizing wastewater treatment processes.

Despite the presented applications of this technology, there are still several problems that need to be solved to fully exploit its potential. The first is the need for relatively frequent calibration of sensors, as they tend to decalibrate themselves. In addition, machine learning models could predict this kind of situation and adjust their operation to the length of time the system is used. Humidity and temperature fluctuations are challenging environmental factors, as they require further improvement to the quality of sensors [173,174] and signal processing algorithms to increase their reliability during normal use. From the organizational side, there is also a lack of standardized norms for sensor validation and interpretation of their indications, which makes it difficult to compare results from tests performed on devices from different manufacturers. Therefore, further research should focus on preparing common guidelines for the requirements for such devices, which would ensure the repeatability of test results and wider application of this technology.

Analyzing the potential directions of development, it can be expected that the next generation of e-noses will be smaller. Their energy efficiency will be improved and their integration with the Internet of Things (IoT) and edge computing will be facilitated. For example, portable e-noses can be used for continuous monitoring of a given person or for analyzing the environment in the workplace, especially in industrial plants. Moreover, the development of federated learning and data analysis in the computing cloud can increase the adaptability and scalability of this kind of system. Furthermore, combining this information with data comes from other sources, like an optical spectroscopy or mass spectrometry, which can help in multidimensional analysis of odors, overcoming current limitations, such as selectivity or cross-sensitivity.

In summary, electronic nose systems are becoming a more and more developed field that can have a major impact on changes in industry and on improving the quality of life. Despite significant progress in improving the sensors themselves and data processing, an interdisciplinary approach to solving problems will be necessary to fully exploit the potential of these technologies. E-nose systems will be increasingly integrated into everyday life, from their use in smart homes and healthcare, to assistance in agriculture or environmental protection. The ability to analyze odors in real time, which will be available, inexpensive, and ubiquitous, will be very important here.