Adaptive E-Nose: Integrating New Gas Sensors for Emerging Applications

Abstract

Conventional chemical analysis relies on costly laboratory instrumentation, while current e-nose systems are expensive for widespread deployment. New opportunities for low-cost, accessible e-nose applications are emerging for diverse fields due to the rapid evolution of inexpensive sensor technologies. We developed a framework that enables rapid integration of newly available low-cost gas sensors into functional e-nose systems, continuously evaluating them as they become commercially available. By characterizing their performance in multi-sensor arrays that mimic biological olfaction, the framework demonstrates effective odor discrimination in a low-cost e-nose system through coordinated behavior of a heterogeneous sensor array. Our testing approach includes sensor sensitivity, selectivity, and stability, which are to be combined with appropriate pattern recognition and AI algorithms in the future for effective chemical discrimination. This work provides a pathway for continuously updating e-nose technology with the latest available sensors in a cost-effective manner, thereby making advanced chemical sensing accessible for resource-limited settings and enabling large-scale deployment in real-world applications with future potential applications such as food quality monitoring, environmental sensing, smart agriculture, etc.

1. Introduction

All life forms sense odors or gather chemicals (e.g., volatile organic compounds (VOCs)) to collect vital information about their surroundings for survival. Terrestrial vertebrates have a nose with a dual role in olfaction (smell) and respiration (breathing) [1,2]; aquatic vertebrates have nostrils (nares) [3]; invertebrates have chemical sensors (chemoreceptors) in their bodies [4], and plants have specialized receptors at a cellular level [5]. Although other senses have been technologically replicated, such as cameras for sight and microphones for hearing, the sense of smell [6] has been difficult to replicate. This is because the biological olfactory system is so sophisticated, being made up of intricate turbinates and millions of receptors to create unique signal patterns for thousands of odorants in milliseconds.

This highly refined biological olfactory function has long driven scientists and engineers to innovate and design systems to emulate it, leading to the development of artificial olfaction systems known as electronic noses (e-noses): systems comprising a variety of non-specific sensors and advanced processors, mimicking nature’s remarkable sensitivity and pattern-recognition capabilities. These systems help us understand and solve problems in various fields of application, such as food spoilage detection and quality control, environmental monitoring, medical diagnostics through breath analysis, and the identification of hazardous substances, including explosives and chemical threats. Although not a physical organ for scent, this concept provides a heightened, often hyper-vigilant, sensitivity to materials deemed offensive, dangerous, or simply inappropriate for public safety. This type of “nose” does not sniff out pheromones but rather sniffs out data from metal oxide sensors, which is the result of a series of chemical reactions. For example, calorimetry sensors respond to thermal changes, whereas optical sensors react to specific light wavelengths. These signals can then be used to detect, identify, or measure target compounds, resulting in a safe environment through effective threat screening and filtering.

Since the conceptual development of a practical sensor array in the 1980s [7] and the subsequent definition of the term “electronic nose” in 1994 [8], the technology has led to the development of e-noses [9] for applications in various domains such as waste management [10], medical applications [11,12,13] and agriculture and forestry [14].

In recent years, innovative technology developments in electronic nose (e-nose) systems have emerged and aim to enhance odor classifications through advanced multivariate analysis and machine learning techniques. Traditional analytical techniques, which are time-consuming and expensive, requiring extensive sample preparation and trained technicians, have given room for the development of various e-noses, such as those based on single-wall carbon nanotubes (SWCNTs) in the wine industry [15], VOC pattern recognition and multivariate signal analysis in the medical industry [16], the detection of volatile organic gases through multivariate analysis in manufacturing industries [17], and machine learning methods for analysis of multidimensional signals in wastewater treatment [18,19].

However, despite innovations and advancements, current electronic nose technology still faces significant challenges, and there is always room for improvement in effectiveness in real-world applications.

(a)

Low Specificity (Poor Selectivity) and Environmental Interferences

The most fundamental challenge is the low specificity (sensors reacting to all compounds rather than only one chemical) of the sensors, especially for the metal oxide (MOS) type. These sensors are broadly cross-sensitive and make it difficult to distinguish between complex mixtures (like coffee aroma or polluted air) that share overlapping volatile organic compounds (VOCs) [20], as the highly ambiguous signal is difficult to deconstruct. This ambiguity is severely worsened by environmental factors, as sensor responses are highly susceptible to changes in ambient humidity and temperature [21], which can alter the sensor’s baseline and sensitivity, leading to inaccurate and highly unreliable readings in real-world conditions. Studies on Fe-doped SnO₂ Acetone Sensors [22] showed how physisorption (physical adsorption) and chemisorption (chemical adsorption) control the gas-sensing process. Physisorption plays a major role at room and lower temperatures, maintaining stability, high sensitivity and strong selectivity toward acetone. As the operating temperature rises, chemisorption becomes more significant, boosting surface reactivity and sensitivity. However, this thermally activated condition can also raise the chance of cross-sensitivity to other gases, which might reduce the sensor’s selectivity.

(b)

Sensor Drift

The sensor is a device made up of active and passive materials to detect physical, chemical, or biological changes in its environment and convert them into output signals that are comprehensible. However, within real-world applications, sensor drift—a slow progressive change—occurs in the output signal without any corresponding changes in the actual quantity being measured. So sensor drift is arguably the most significant barrier to the long-term reliability of electronic nose technology [23]. This phenomenon is caused by physical and chemical degradation (aging) of the sensing material (e.g., the metal oxide layer) and “poisoning” from irreversible chemical binding. This drift means that the “fingerprint” pattern a sensor array produces for a specific odor today will be different from the pattern it produces weeks or months later, invalidating the trained machine learning model and forcing frequent, time-consuming recalibrations of the entire system. Sensor drift is limited by various factors such as extreme temperature fluctuations, constant mechanical stress, electrical noise and interference, sensor component aging, and sensor surface contamination and corrosion. Sensor drift can be subtle, but there might be an impactful issue that undermines the very purpose of measurement. So sensor-drift compensation [24,25,26] is necessary to ensure long-term reliability and accuracy. However, this aspect of sensors is not covered within this study; we currently focus on establishing a proof-of-concept and demonstrating the feasibility of the proposed heterogeneous sensor array framework across multiple applications.

(c)

Bulkiness and High-Power Consumption

While e-nose systems have evolved toward greater portability, their inefficient system design makes them unsuitable for flexible field deployment. Even though the core components, such as gas sensor arrays, transmission paths, microprocessors, and pattern recognition algorithms [9], are necessary for operation, the true portability and accessibility are compromised by their conventional integration. More importantly, despite their superior sensing capabilities, the widely used MOS sensors [27,28,29]—the Figaro Series, such as TGS2600 and TGS3870; Winsen Series, such as MQ-2, MQ-8, MQ-135, and Alphasense Series, such as MiCS-6814—have historically required integrated heaters that operate at 100–500 °C [30], consuming significant power that prevents long-term battery operation. Due to this design limitation, current e-nose systems can only be used in laboratory or stationary industrial environments, which restricts their use in dynamic field situations where compact, energy-efficient operation is crucial.

The shortfalls mentioned above in numerous electronic nose systems have motivated us to develop a novel approach to developing a framework to test and validate new sensors coming into the market; new flexible algorithms are made available, providing optimum portability for varied applications. Researchers have conducted studies on various substances, such as apple [31,32,33], mango ginger [34], etc., but only on the existing, established e-nose system. In this study, we have designed a novel system using low-cost gas sensors that will provide optimum analysis of odor signatures and a deep understanding of how they react to various aroma characteristics, capturing a unique gas signature for each substance. It can be used for diverse applications as required, ranging from the detection of explosives, the monitoring of food spoilage, water quality, and the environment, to the identification of hazardous substances.

2. Materials and Methods

2.1. Overview of Sensing Array

The array of sensors in an electronic nose system is its core component, acting as an “artificial olfactory receptor” for the device. These sensors are designed to mimic the non-specific yet comprehensive sensing ability, as that of the biological nose, in recognizing and classifying odors. Typically, an array contains different kinds of chemical sensors: homogeneous (same type, or slight variations) or heterogeneous (different types) as in

Table 1. Sensor types.

Sensor Type	Operating Principle	Sensing Material
MOS Gas Sensor (Metal Oxide Semiconductor)	Chemiresistive—target gas adsorbs on heated SnO2 surface, causing a resistance change	Tin Dioxide (SnO2) on alumina substrate, noble-metal electrodes
MEMS MOX Gas Sensor (Multi-pixel)	Chemiresistive—gas reacts on MOX micro-hotplate; digital I2C output with onboard signal processing	Metal Oxide (MOX) on CMOS micro-hotplate with multiple independent sensing pixels
Electrochemical Gas Sensor (Amperometric)	Amperometric—target gas is oxidized at the working electrode, generating a current proportional to concentration; auxiliary electrode corrects zero drift	Platinum (Pt) and Carbon (C) catalyst electrodes in aqueous electrolyte with PTFE diffusion membrane
NDIR Gas Sensor	Infrared absorption at the CO2-specific wavelength	Infrared optical cavity (IR Source + detector pair)

.

Based on the applications, the sensors in the array collect electrical signals in a multidimensional response pattern, which can be analyzed by computer-system algorithms trained using a database of known odor “fingerprints.” The new odor pattern is compared with the patterns in the library and is then classified and identified as the closest match, effectively translating the unique fingerprint into a recognized smell (e.g., fruit, contaminant, etc.).

2.2. Experimental System Setup

In this experimental system setup (Figure 1), we have used an array of sensors that are exposed to various kinds of gases. The Metal Oxide Semiconductor (MOS) sensors, such as TGS2600 (Figaro Engineering Inc., Minoh, Japan), TGS2602 (Figaro Engineering Inc., Minoh, Japan), TGS2603 (Figaro Engineering Inc., Minoh, Japan), BME688 (Seeed Studio Technologies, Shenzhen, China), and SGP30 (Seeed Studio Technologies, Shenzhen, China), Multichannel Sensor (Seeed Studio Technologies, Shenzhen, China), and electrochemical sensor, such as SFA30 (Seeed Studio Technologies, Shenzhen, China), are grouped and connected in one board, as in

Table A1. Sensors connected to Seeeduino Cortex-M0+ Board 1 (adapted from [51,52]).

Sensor Module—Make	Sensing Technology	Target Gases	Measurement Range	Cross Sensitivity
TGS2600-B00—Figaro Engineering Inc.	Metal Oxide Semiconductor (SnO2)	Air contaminants (H2, CO, ethanol, methane)	1–30 ppm H2	reducing gases, VOCs
TGS2602-B00—Figaro Engineering Inc.	Metal Oxide Semiconductor (SnO2)	Odorous Gases (NH3, H2S, etc.), VOCs	1–30 ppm ethanol	n/a
TGS2603—Figaro Engineering Inc.	Metal Oxide Semiconductor (SnO2)	Air Contaminants (Trimethylamine, methyl mercaptan, H2S, ethanol, etc.)	1–10 ppm H2	n/a
Grove—Air Quality Sensor (BME688)—Seeed Studio with a chip from Bosch Sensortec	Metal Oxide Semiconductor (MOS) with Artificial Intelligence (AI)	Volatile Organic Compounds (VOCs), Volatile Sulfur Compounds (VSCs), Carbon Monoxide (CO), and Hydrogen (H2)	Parts per billion (ppb) range for target gases	n/a
Grove VOC and eCO2 Gas Sensor (SGP30) v1.0	Seeed Studio with a chip from Sensirion—Metal Oxide Semiconductor (MOS)	TVOCs, eCO2	TVOC: 0–60,000 ppb & eCO2: 400–60,000 ppm	n/a
Multichannel Gas Sensor v2.0—Seeed Studio with chips from Winsen	Metal Oxide Semiconductor (MOS) with Micro-Electro-Mechanical System (MEMS)	Nitrogen dioxide (NO2), Ethyl Alcohol (C2H5OH), VOC, CO	NO2: 0.1–10 ppm & C2H5OH: 1–500 ppm & VOC: 1–500 ppm & CO: 5–5000 ppm	n/a
Grove—Formaldehyde Sensor (SFA30) v1.0—Seeed Studio with a chip from Sensirion	Electrochemical	Formaldehyde (HCHO) with integrated RTH (Relative Humidity & Temperature)	0–1000 ppb	n/a

, while other MOx sensors, such as TGS2610 (Figaro Engineering Inc., Minoh, Japan), TGS2611 (Figaro Engineering Inc., Minoh, Japan), TGS2612 (Figaro Engineering Inc., Minoh, Japan), BME680 (Seeed Studio Technologies, Shenzhen, China), SGP41(Seeed Studio Technologies, Shenzhen, China) and Non-Dispersive Infrared (NDIR) CO₂ gas Sensors (Seeed Studio Technologies, Shenzhen, China) are placed together in another board, as in

Table A2. Sensors connected to Seeeduino Cortex-M0+ Board 2 (adapted from [51,52]).

Sensor Module/Make	Sensing Technology	Target Gases	Measurement Range	Cross Sensitivity
TGS2610—Figaro Engineering Inc.	Metal Oxide Semiconductor (SnO2)	LPG (Liquefied Petroleum Gas)—mixture of Propane (C3H8) and Butane (C4H10)	500–10,000 ppm (or 1–25% Lower Explosive Limit (LEL))	Ethylene, Alcohols, organic vapours
TGS2611—Figaro Engineering Inc.	Metal Oxide Semiconductor (SnO2)	Methane (CH4) and natural gas	500–10,000 ppm (or 1–25% Lower Explosive Limit (LEL))	low sensitivity to Hydrogen (H2), Iso-butane (i-C4H10), Ethanol (C2H5OH)
TGS2612—Figaro Engineering Inc.	Metal Oxide Semiconductor (SnO2)	Methane (CH4), Propane (C3H8), Iso-Butane (i-C4H10) (components of LNG/LPG)	1–25% Lower Explosive Limit (LEL) of each gas	Hydrogen (H2), Ethanol (C2H5OH)
Grove—Barometer Sensor (BME680) v1.0—Seeed Studio with a chip from Bosch Sensortec	Metal Oxide Semiconductor (MOS) with MEMS-based for Temperature, Pressure & Humidity	Volatile Organic Compounds (VOCs), temperature, humidity, and air pressure	Temp.: −40–+85 °C, Humidity: 0–100% r.H, Pressure: 300–1100 hPa, VOC: Broad range, typically measured in kΩ	Hydrogen (H2)—moderate, Carbon monoxide—limited
Grove—VOC and NOx Gas Sensor (SGP41) v1.0—Seeed Studio with a chip from Sensirion	Metal Oxide Semiconductor (MOS) utilizing CMOSens® and MOXSens® technologies	VOC Pixel (Reducing Gases): VOCs, NOx Pixel (Oxidizing Gases): Nitrogen Dioxides (NO2)	VOCs: 0–1,000,000 ppb (Ethanol in clean air), VOC index: 0–500, NOx: 0–10,000 ppb (NO2 in clean air), NOx index: 0–500	Hydrogen (H2), Ozone (O3)
Grove—CO2 Sensor (MH-Z16)—Seeed Studio	Non-Dispersive Infrared (NDIR)	Carbon Dioxide (CO2)	0–2000 ppm	Highly selective for CO2; negligible cross-sensitivity to other gases

.

Additionally, to provide more diversity to the sensing process, eight calibrated Libelium sensors (Libelium Comunicaciones Distribuidas, Zaragoza, Aragon, Spain) (

Table A3. Sensors mounted on Waspmote Gases PRO Sensor Board-1 (adapted from [53]).

Sensor Module/Make	Sensing Technology	Target Gases	Measurement Range	Cross Sensitivity
Ammonia Gas Sensor (4-NH3-100)—Honeywell Analytics, Lincolnshire, IL, USA	Electrochemical principle	Ammonia (NH3)	0–100 ppm	0 ppm@300 ppm CO, 1.5 ppm@5 ppm H2S,−3 ppm@5 ppm CO2,30 ppm@15 ppm H2,−1 ppm@35 ppm Isobutylene,0@100 ppm Ethanol
Nitric Dioxide Gas Sensor (NO2-A43F)—Alphasense Ltd., Braintree, UK	Electrochemical principle	Nitrogen Dioxide (NO2)	0–20 ppm	<−80 ppm@5 ppm H2S,<5 ppm@5 ppm NO,<75 ppm@5 ppm Cl2,<−5 ppm@5 ppm SO2,<−5 ppm@25 ppm CO,<−0.1 ppm@100 ppm H2,<1 ppm@100 ppm C2H4,<0.2 ppm@20 ppm NH3, <0.1 ppm@5% vol CO2,nd @ 100 ppm Halothane
Nitric Oxide Gas Sensor (NO-A4)—Alphasense Ltd., Braintree, UK	Electrochemical principle	Nitric Oxide (NO)	0–18 ppm	<0 ppm@300 ppm CO,<0 ppm@5 ppm SO2,<1.5 ppm@5 ppm NO2,<−1.5 ppm@15 ppm H2
Carbon Monoxide Gas Sensor (CO-A4)—Alphasense Ltd., Braintree, UK	Electrochemical principle	Carbon Monoxide (CO)	0–25 ppm	<0.1 ppm@5 ppm H2S,<−2 ppm@5 ppm NO2,<0.1 ppm@5 ppm Cl2,<−2 ppm@5 ppm NO,<0.1 ppm@5 ppm SO2,<10 ppm@100 ppm H2,<0.5 ppm@100 ppm C2H4,<0.1 ppm@20 ppm NH3

and

Table A4. Sensors mounted on Waspmote Gases PRO Sensor Board-2 (adapted from [53]).

Sensor Module/Make	Sensing Technology	Target Gases	Measurement Range	Cross Sensitivity
Methane and Combustible Gas Sensor (CH-A3)—Alphasense Ltd., Braintree, UK	Catalytic combustion (also known as pellistor technology)	Methane (CH4)	0–100% Lower Explosive Limit (LEL) Methane	160–175%LEL@130–140% Sensitivity Hydrogen, 350–450%LEL@150–190% Sensitivity Propane, 420–500%LEL@150–180% Sensitivity Butane, 600–670%LEL@180–200% Sensitivity n-Pentane, 800–950%LEL@150–170% Sensitivity Nonane, 17–18%LEL@42–44% Sensitivity Carbon Monoxide, 300–340%LEL@150–170% Sensitivity Acetylene, 270–320%LEL@150–170% Sensitivity Ethylene, 450–500%LEL@180–200% Sensitivity Isobutylene
Ozone Gas Sensor (OX-A431)—Alphasense Ltd., Braintree, UK	Electrochemical principle	Ozone (O3)	0–18 ppm	<100 ppm@5 ppm H2S, <70–120 ppm@5 ppm NO2, <30 ppm@5 ppm Cl2, <3 ppm@5 ppm NO, <−6 ppm@5 ppm SO2,<0.1 ppm@5 ppm CO,
Ozone Gas Sensor (OX-A431)—Alphasense Ltd., Braintree, UK	Electrochemical principle	Ozone (O3)	0–18 ppm	<0.1 ppm@100 ppm H2, <0.1 ppm@100 ppm C2H4, <0.1 ppm@20 ppm NH3, <0.1 ppm@50,000 ppm CO2, <0.1 ppm@100Halothane
Sulfur Dioxide Gas Sensor (SO2-A4)—Alphasense Ltd., Braintree, UK	Electrochemical principle	Sulfur Dioxide (SO2)	0–20 ppm	<40 ppm@5 ppm H2S, <−160 ppm@5 ppm NO, <−70 ppm@5 ppm Cl2, <−1.5 ppm@5 ppm SO2, <2 ppm@5 ppm CO, <1 ppm@100 ppm H2, <1 ppm@100 ppm C2H4, <0.1 ppm@20 ppm NH3, <0.1 ppm@5% vol CO2
LF02-A4—Alphasense Ltd., Braintree, UK	Electrochemical principle	Oxygen (O2)	0–30% Vol.%	n/a
Hydrogen Sulfide Gas Sensor (4-H2S-100)—Honeywell Analytics, Lincolnshire, IL USA	Electrochemical principle	Hydrogen Sulfide (H2S)	0–100 ppm	≤6 ppm@50 ppm CO, 1 ppm@5 ppm H2S, 1 ppm@35 ppm NO, −1 ppm@5 ppm NO2, 25 ppm@10,000 ppm H2, <0 ppm@100 ppm C2H4, ±1.5 ppm@5000 ppm C2H6O

) mounted on two Waspmote Gases PRO Sensor Boards (Libelium Comunicaciones Distribuidas, Zaragoza, Spain), featuring the ATmega1281 microcontroller, have been integrated into the system. To each board, a set of batteries is attached to power the Real-Time Clock (RTC).

They are all housed in a 5.4 L inner capacity test chamber of size 235W × 180D × 210H (mm), fitted with a small mixing fan to circulate the inner air uniformly. The array of sensors that are connected to the Seeeduino boards (Seeed Studio Technologies, Shenzhen, China) has been installed in the upper chamber, and Libelium sensors mounted on the Waspmote Gases PRO sensor board are installed in the lower chamber. The chamber also has an air filter attached to maintain the internal environment optimally. An external 5 VDC power supply is supplied to all boards to stabilize the sensor readings.

2.3. Sensors and Specifications

The sensors are an integral part of the electronic nose system, and an array of them would constitute multiple gas sensors that respond to various gases. The different sensor technologies [35] capture a unique fingerprint produced by gases. Existing pattern recognition algorithms [36] can be used to identify and quantify the gases with more accuracy, despite the cross-sensitivity of the individual sensor. In this experimental setup, we have included the following sensors (in

Table 2. Sensors and sensing technologies.

Sensors	Sensing Technology
TGS2600, TGS2602, TGS2603, TGS2610, TGS2611, TGS2612	Metal Oxide Semiconductor (SnO2)
Air Quality Sensor (BME688)
VOC and eCO2 Gas Sensor (SGP30) v1.0	Metal Oxide Semiconductor (MOS)
Multichannel Gas Sensor v2.0	Metal Oxide Semiconductor (MOS) with Micro-Electro-Mechanical System (MEMS)
Formaldehyde Sensor (SFA30) v1.0	Electrochemical
Barometer Sensor (BME680) v1.0	Metal Oxide Semiconductor (MOS) with MEMS-based for Temperature, Pressure & Humidity
VOC and NOx Gas Sensor (SGP41) v1.0	Metal Oxide Semiconductor (MOS) utilizing CMOSens® and MOXSens® technologies
CO2 Sensor (MH-Z16)	Non-Dispersive Infrared (NDIR)
Ammonia Gas Sensor (4-NH3-100)	Electrochemical principle
Nitric Dioxide Gas Sensor (NO2-A43F)	Electrochemical principle
Nitric Oxide Gas Sensor (NO-A4)	Electrochemical principle
Carbon Monoxide Gas Sensor (CO-A4)	Electrochemical principle
Methane and Combustible Gas Sensor (CH-A3)	Catalytic combustion (also known as pellistor technology)
Ozone Gas Sensor (OX-A431)	Electrochemical principle
Sulfur Dioxide Gas Sensor (SO2-A4)	Electrochemical principle
Oxygen Gas Sensor (LF02-A4)	Electrochemical principle
Hydrogen Sulfide Gas Sensor (4-H2S-100)	Electrochemical principle

) and organized them into their respective development board groups.

Detailed specifications for every sensor mounted on the four acquisition boards—including sensing technology, target gases, measurement range, and cross-sensitivity—together with the individual sensor descriptions, are provided in Appendix A (Table A1, Table A2, Table A3 and Table A4).

2.4. Fruits and Their Scent

Fruit scent determines the quality of the fruit, providing the consumer with the best indicator of fruit flavor. The wide variety of volatile organic compounds (VOCs) released by fruits (in

Table 3. Fruit samples and their aroma/scents.

Fruit Type	Ripening Compound	Aroma Compound	Other Gases	Reference
Apple	Ethylene (ethene, C2H4)	VOCs (esters, alcohols, ketones, alkenes, aldehydes, acid, phenol)	Carbon Dioxide (CO2)	[32,33]
Banana	Ethylene (ethene, C2H4)	VOCs (esters, alcohols, ketones, alkenes, aldehydes, acid, phenol, hydrocarbons)	Carbon Dioxide (CO2)	[38]
Mango	Ethylene (ethene, C2H4)	VOCs (esters, alcohols, ketones, alkenes, aldehydes, acid)	Carbon Dioxide (CO2)	[34,39]
Orange	Ethylene (ethene, C2H4)	VOCs (esters, alcohols, ketones, alkenes, aldehydes, acid)	Carbon Dioxide (CO2)	[40]
Strawberry	Ethylene (ethene, C2H4)	VOCs (terpenes, esters, aldehydes, lactones, ketones, acids)	Carbon Dioxide (CO2)	[41,42]
Pears	Ethylene (ethene, C2H4)	VOCs (esters, aldehydes, alcohols, ketones, alkenes, acids)	Carbon Dioxide (CO2) and Nitrogen (N2)	[43]

) is studied to identify, characterize, and grade different fruits [37]. The fruits usually produce chemical compounds that help in either ripening or contributing to aroma. The following fruits (Figure 2) have been tested for their contained compounds, as observed depending on their respective cultivar, harvesting, aging, etc.

2.5. Software System and Data Acquisition

Sensor data were acquired, stored, and visualized in real time by a modular software stack. Python (v3.14.6) acquisition scripts running on a central computer stream readings from the four sensor boards, save them to local CSV files, and upload them to a Firebase Realtime Database cloud backend, while a browser-based Plotly (v6.8.0) Dash dashboard provides live monitoring and experiment control.

2.5.1. Architecture Overview

The e-nose software system architecture, as seen in Figure 3, has been developed to reliably collect, store, and visualize sensor data in real time during laboratory experiments. It connects four sensor nodes—two custom gas-sensing boards (B1 and B2) and two Libelium environmental boards (LB1 and LB2)—to a central acquisition computer that controls the experiment workflow.

Python scripts running on the acquisition computer handle serial communication with all sensor boards, process incoming data, save readings to local CSV files, and upload the same information to a cloud database. A web-based dashboard built using the Dash framework allows researchers to configure experiments and observe sensor responses live through a standard web browser.

This layered design allows the system to remain modular and easy to extend. For example, new sensors can be added without changing the overall architecture, and the dashboard can be accessed remotely without installing additional software.

2.5.2. Data Acquisition Workflow

The acquisition engine is written in Python and is responsible for all communication between the sensor hardware and the software system. A central script, write.py, controls the entire acquisition process. The sequence of data flow in the system is as shown in Figure 4. It reads user-defined experiment parameters from the dashboard, creates the appropriate directory structure, and starts the board-specific data listeners. Separate scripts handle different board types: one for the high-frequency gas sensors (B1 and B2) and another for the environmental sensors (LB1 and LB2). These scripts run in parallel threads so that data collection is not interrupted when one board responds more slowly than others. This modular approach ensures stable data collection even when the system is under heavy load or when individual boards behave unpredictably.

Further implementation details—the dashboard interface, the data collection and synchronization strategy, Firebase integration, and the real-time analytics—are provided in Appendix B.

2.6. Experimental Process

The experiment was conducted by initially fanning out any contaminants inside the room and the test chamber. After it is ensured that the setup is free of unwanted gases, the test chamber is closed without any samples inside it. The 220 VAC power supply is connected to the test chamber to run the mixing fan inside, and a 5 VDC supply to power up all the boards. The system is then kept running for 30 min to stabilize the sensor readings.

The baseline measurements are captured and recorded for 20 min. The test chamber is opened, and a piece of paper is placed inside as a base for the substance sample so that no aroma or scent is left on the floor of the chamber to avoid cross-contamination of odor. After each test is completed, it is replaced with a new paper to maintain the purity of the air inside the chamber. A substance sample is placed on the paper base, and the internal mixing fan evenly distributes accumulated gases. The process is kept running for 20 min, and data is captured and recorded both locally in a CSV file and uploaded in real time to the Firebase cloud. Once the data recording for a specific substance (e.g., fruit) is completed, the test chamber is opened to let out any gases accumulated inside the chamber. The test process was diligently repeated for all other substances. For the current set of experiments, we focused on the gases released by fruits, and no specific gas was injected. However, for future experiments, in which actual gases are to be introduced into the chamber, a hole has been provisioned at the bottom, as seen in Figure 1c. The sequence of the experimental process conducted in a laboratory setup is shown in Figure 5 below.

2.7. Tools for Statistical Analysis

All statistical analyses were conducted using IBM SPSS Statistics Version 30. Parametric statistical methods were applied depending on data distribution characteristics. Normality of the sensor response datasets was assessed using the Shapiro–Wilk test, which guided the selection of subsequent analyses. For normally distributed data, one-way analysis of variance (ANOVA) was performed, followed by Tukey–Kramer post hoc comparisons to identify statistically significant pairwise differences between fruit samples.

The recorded sensor responses represent time series measurements obtained during controlled exposure experiments, rather than independent biological replicates. Each fruit sample was measured over a defined exposure period, generating 52 temporal observations under identical conditions. These time-resolved data were used to characterize sensor response dynamics; however, they are not treated as statistically independent replicates in the strict experimental sense. Instead, the analysis focuses on identifying consistent and reproducible response patterns across substances, supported by both univariate (ANOVA) and multivariate methods.

In addition to univariate analysis, multivariate analysis was conducted using Principal Component Analysis (PCA) applied to standardized sensor-response features. PCA was used to reduce dimensionality, identify dominant variance patterns, and evaluate the overall structure of the dataset. This approach enabled visualization of clustering behavior and assessment of the ability of the e-nose system to discriminate between fruit types.

3. Results and Discussions

Existing studies have shown that ethylene is either released by fruits naturally or is used commercially to alleviate the ripening process [44,45]. As the ripening process increases, carbon dioxide [46] is released as a normal metabolic product, along with water vapor [47], which would raise humidity in a closed container. Additionally, many VOCs [48,49], such as esters, alcohols, aldehydes and terpenes, are released by fruits, which give the characteristic smell of each fruit. The VOCs are typically present at low concentrations compared with carbon dioxide and ethylene, but they are important for flavor and perceived freshness [47]. For this reason, many studies [32,34,38,39,43] have been conducted on VOCs released from fruits, and the output of these studies by other authors has helped us in better understanding the gases released by the fruits used in our experiment. So, in this study, out of the many gases detected, we have focused on the sensor’s responses to the VOC and CO₂ equivalent, in accordance with our main objective of examining the selectivity and sensitivity of low-cost sensors for emerging applications.

3.1. Statistical Analysis

3.1.1. Grove—Air Quality Sensor (BME688) Response to VOC and Carbon Dioxide (CO₂) Equivalent Gas

The boxplots in Figure 6a,b below show that the air quality sensor (BME688) clearly differentiates between the baseline background and all tested fruits using both VOC and CO₂ equivalent outputs. Statistical analysis (one-way ANOVA, F = 4522, df = 6, p < 0.001; Dunnett T3, 95% CI) provides strong evidence that the VOC responses are significantly different across fruit types. The BME688 VOCs metal oxide-sensing element exhibits known cross-sensitivity to a wide range of volatile compounds released by fruits, enabling effective discrimination based on their vapor signatures.

A similar trend is evident in the CO₂ equivalent channel (one-way ANOVA, F = 19,693, df = 6, p < 0.001; Dunnett T3, 95% CI). Despite not directly detecting CO₂ from the fruits, the sensor interprets reducing VOCs as CO₂ equivalent signals due to its chemical response mechanism. As a result, the CO₂ equivalent output also successfully separates fruit samples from the baseline air.

Across both channels, all fruit types produce significantly higher responses than the baseline, with distinct median values and minimal overlap. Strawberry and mango generate the strongest responses, consistent with their high emission of volatile aromatic compounds. Banana and orange show intermediate responses, while apple and pear produce lower signals but remain clearly distinguishable from the baseline. Although fruits are not sources of CO₂, the elevated CO₂ equivalent readings reflect the sensor’s cross-sensitivity to reducing VOCs. This effect can be advantageously exploited for odor-based classification.

As illustrated in Figure 6c, the VOC equivalent and CO₂ equivalent responses separate all fruit types with high statistical confidence (ANOVA: F = 19,693, df = 6, p < 0.001; Dunnett T3, 95% CI). Although BME688 does not directly detect CO₂, its MOS-based gas-sensing mechanism interprets reducing VOCs as CO₂ equivalent signals—a behavior that enhances pattern recognition capability for VOC-rich substances such as fruits.

Strawberry and mango samples consistently produce the strongest responses, reflecting their complex VOC emissions, while banana and orange show intermediate patterns. Apple and pear display lower signals but remain clearly differentiated from the baseline. The particularly low response from pear samples may relate to limited VOC release at the measured ripeness stage or a VOC profile that partially falls outside the sensor’s optimal sensitivity range.

3.1.2. Grove—Air Quality Sensor (SGP30) Response to Carbon Dioxide (CO₂) Equivalent Gas

Figure 7 shows the SGP30 sensor’s response to measuring CO₂ equivalent in ppm across different fruit samples. The sensor has an exceptionally high response to strawberries, while the responses from other fruit samples are close to the baseline measurement. This is because SGP30 is a metal oxide (MOx) gas sensor that uses multiple sensing elements to detect volatile organic compounds (VOCs) and calculates CO₂ equivalent values based on the total VOC concentration (Figure 8a) it detects, rather than actual CO₂. Since strawberries are rich in distinct VOCs that give them a unique and abundant aromatic compound profile, the sensor shows high cross-sensitivity to those compounds, and it measures CO₂ equivalent by detecting and quantifying the total VOC load, which is significantly higher in strawberries.

3.1.3. Responses of Various Sensors to VOCs

Previous studies employing GC-MS technology have consistently shown that strawberries emit a complex mixture of volatile organic compounds (VOCs) [50], including esters, alcohols, aldehydes, terpenes, and lactones, with concentrations strongly influenced by cultivar, ripeness stage, and post-harvest handling conditions. These VOCs form the strawberry’s characteristic aroma signature and provide a useful benchmark for assessing sensor performance. To determine which of our sensing technologies are most effective in capturing these signatures, strawberries were placed inside a controlled test chamber and evaluated using multiple gas sensor platforms.

Figure 6a and Figure 8a demonstrate that both the BME688 and SGP30 sensors detect markedly elevated VOC levels across all tested fruits, with especially pronounced responses for strawberries. Statistical analysis for Figure 8a confirms significant differences between fruit types (ANOVA: F = 3866; df = 6; p < 0.001; Dunnett T3, 95% CI). These results align with manufacturer-published sensitivity profiles for BME688 and SGP30, which highlight their strong responsiveness to reduced organic vapors and aroma-active VOCs. The narrow inter-quartile ranges observed in these boxplots further indicate stable, consistent VOC measurements across replicate strawberry samples. This reliability underscores the suitability of these sensors for applications such as strawberry ripeness assessment, quality grading, and early spoilage detection, which similarly rely on subtle variations in VOC composition.

The Multichannel MEMS-based gas sensor (Figure 8b) exhibits mid-range VOC detection performance, with statistically significant differences between fruit types (ANOVA: F = 1975; df = 6; p < 0.001; Dunnett T3, 95% CI). However, its responses show greater variability than those of BME688 and SGP30. This wider dispersion likely reflects its multi-element design, through which each sub-sensor responds to different classes of VOCs. Such variation has been documented in earlier studies, where mechanical impact, storage temperature, and cultivar-specific volatiles produced heterogeneous VOC profiles in strawberries and other fruits. While this variation may reduce precision for ripeness scoring, it may be advantageous for broader chemical fingerprinting applications.

In contrast, SGP41 (Figure 8c) employs Sensirion’s advanced CMOSens® and MOXSens® technologies, which compute a combined VOC index ranging from 0–500. The observed readings near VOC Index ≈ 300 fall within an optimal detection zone, indicating strong sensitivity to strawberry-associated reducing VOCs and a robust response window well-suited for distinguishing high-aroma fruits from background levels. Statistical analysis supports this (ANOVA: F = 4869; df = 6; p < 0.001; Dunnett T3, 95% CI). Previous work evaluating CMOSens-based sensors similarly reports stable behavior and reduced noise when detecting mixed VOC environments, making them promising candidates for monitoring fruit freshness in dynamic, real-world conditions such as storage rooms or transport containers.

3.2. Multivariate Analysis of Sensor Responses Using Principal Component Analysis

While individual sensor responses provide valuable insight into sensitivity and selectivity, electronic nose systems fundamentally rely on multivariate pattern recognition to discriminate complex odor mixtures. To evaluate the collective behavior of the heterogeneous sensor array and to identify the sensing elements contributing most significantly to discrimination, principal component analysis (PCA) was performed using IBM SPSS Statistics version 30 on the baseline-subtracted dataset. Substance 0 (background air) was used only to establish the baseline and was excluded from the PCA model for the six fruits: strawberry (1), banana (2), orange (3), mango (4), apple (5), and pear (6).

Prior to principal component analysis, all sensor channels were standardized to zero mean and unit variance by performing PCA on the correlation matrix (See Appendix C, Figure A2). This scaling procedure ensured that differences in measurement units, dynamic ranges, and sensor technologies did not bias the multivariate analysis, allowing each sensor to contribute equally to component extraction.

3.2.1. Sensor Contribution and Communality Analysis

The communalities table (

Table 4. Communalities of sensor variables extracted from principal component analysis (PCA) performed in SPSS 30, indicating the proportion of variance in each sensor, explained by the retained components.

Communalities	Extraction	Communalities	Extraction	Communalities	Extraction
B1—BME688 bVOC Equivalent (ppm)	0.893	B1—IAQ Index	0.835	B2—TGS2611	0.916
B1—BME688 CO Equivalent (ppm)	0.695	B1—Static IAQ	0.695	B2—TGS2612	0.599
B1—SGP30 CO2eq (ppm)	0.855	B1—TGS2600	0.532	B2—SGP 41 VOC Index	0.872
B1—Multi cha2 GM102B (NO2) (ppm)	0.950	B1—TGS2602	0.796	LB1—NO (ppm)	0.934
B1—Multi cha2 GM302B (C2H5CH) (ppm)	0.894	B1—TGS2603	0.611	LB1—CO (ppm)	0.314
B1—Multi cha2 GM502B (VOC) (ppm)	0.949	B1—SGP30 tVOC (ppb)	0.537	LB2—SO2 (ppm)	0.790
B1—Multi cha2 GM702B (CO) (ppm)	0.925	B2—CO2 Concentration (ppm)	0.583	LB2—H2S (ppm)	0.929
B1—Formaldehyde hcho (ppb)	0.733	B2—NOx Raw	0.972	LB2—OX-A431 O3 (ppm)	0.728
		B2—TGS2610	0.013	LB2—O2 (ppm)	0.310

Extraction method: Principal component analysis.

) obtained from the SPSS PCA output presents the proportion of variance in each sensor variable, explained by the retained principal components. Sensors exhibiting high extracted communality values contribute strongly to the multivariate structure of the dataset and, consequently, to odor discrimination.

The highest communalities were observed for:

• B2—NOx Raw (0.972);
• B1—Multi channel GM102B (NO₂) (0.950);
• B1—Multi channel GM502B (VOC) (0.949);
• LB1—NO (0.934);
• LB2—H₂S (0.929);
• B1—Multi channel GM702B (CO) (0.925);
• B2—TGS2611 (0.916);
• B1—SGP41 VOC Index (0.872);
• B1—SGP30 CO₂ equivalent (0.855).

These results demonstrate that both broadly responsive VOC sensors and more selective electrochemical gas sensors play a dominant role in shaping the PCA solution. The strong contribution of BME688, SGP30, and SGP41 confirms the importance of MOS-based VOC detection for odor intensity encoding, while the high communalities associated with NOx, NO, H₂S, and NO₂ sensors indicate that oxidizing- and nitrogen-based gas responses provide critical complementary information.

In contrast, B2—TGS2610 (communalities = 0.013) and LB2—O₂ (0.310) exhibited minimal contribution, suggesting limited relevance for fruit aroma discrimination under the present experimental conditions. This finding provides valuable guidance for future sensor selection and array optimization, particularly for low-power or application-specific deployments.

3.2.2. Eigenvalue Analysis and Component Retention

To determine the appropriate number of principal components, eigenvalue analysis and inspection of the scree plot were performed, as shown in Figure 9. The scree plot reveals a pronounced drop in eigenvalue magnitude between the first and second components, followed by a gradual flattening beyond the second component. The first two principal components (PC1 and PC2) explained approximately 72.48% of the total variance in the dataset, with PC1 accounting for 56.01% and PC2 for 16.47%.

The first principal component (PC1) possesses a substantially larger eigenvalue, capturing the dominant variance in the dataset (56.01%), while the second principal component (PC2) retains an eigenvalue above unity and accounts for a meaningful proportion of the remaining variance (16.47%). Subsequent components exhibit eigenvalues close to zero and contribute only marginal additional information.

Based on the Kaiser criterion (eigenvalues > 1) and the clear inflection point observed in the scree plot, two principal components were retained. Together, PC1 and PC2 explain 72.48% of the total variance, indicating that the majority of the multivariate information contained in the full sensor array can be represented in a two-dimensional feature space.

3.2.3. PCA Scores Plot and Fruit Discrimination

The regression-based PCA scores plot for PC1 (56.01%) and PC2 (16.47%) is shown in Figure 10. Each point represents an individual observation projected into the reduced PCA space, with samples grouped by fruit type: strawberry, banana, orange, mango, apple, and pear.

Clear and compact clustering is observed for all fruit samples, demonstrating that the electronic nose array captures distinct and reproducible odor fingerprints. Separation is dominated by PC1, which reflects overall VOC intensity and reducing gas activity. Fruits with richer and more complex volatile emissions, specifically strawberry and mango, occupy extreme regions along PC1, consistent with their known ester- and terpene-rich aroma profiles.

Banana and orange samples form intermediate clusters, reflecting moderate VOC emission dominated by characteristic esters and citrus terpenes, respectively. Apple and pear cluster closer together at lower PC1 values, indicating weaker overall VOC intensity and chemically similar volatile profiles at the measured ripeness stage.

PC2 provides secondary discrimination, separating fruits with comparable total VOC levels but differing gas composition. This axis is strongly influenced by oxidizing- and nitrogen-based sensor responses, as evidenced by the high communalities of NOx-related sensors. As a result, fruits with subtle differences in VOC composition and oxidation byproducts become separable in the PC1–PC2 plane, even when PC1 intensities overlap.

3.2.4. Qualitative Sensor Contribution Analysis Based on Rotated Component Loadings

The rotated component loading plot (Figure 11) shows how individual sensor variables contribute to the principal components and presents the structure of the multivariate responses. After Varimax rotation, distinct sensor groupings emerge in the PC1–PC2 space.

Sensors with strong positive loadings on PC1, including BME688 (bVOC, IAQ indices, and CO equivalent), SGP30 (CO₂eq and tVOC), SGP41 (VOC index), and MOS sensors, such as TGS2600, TGS2602, and TGS2611, form a compact cluster. Their close proximity indicates high intercorrelation and confirms that PC1 represents overall VOC intensity and reducing gas activity, driving the primary discrimination between fruit samples.

In contrast, sensors such as Multichannel GM102BN (NO_x Raw), and OXA431 (O₃) appear on the negative side of PC1, indicating an inverse relationship with VOC-dominated signals and representing background or oxidizing gas behavior.

The vertical separation along PC2 reflects compositional differences rather than intensity. VOC-related sensors (e.g., SGP30 tVOC and SFA30 formaldehyde) lie in the positive PC2 region, while NO_x-related, SO₂ and oxidizing sensors appear negative, indicating that PC2 captures the balance between reducing and oxidizing gases.

MEMS-based sensors (GM302B and GM502B) occupy intermediate regions, highlighting their role in detecting specific VOC classes, complementing the broader response of MOS sensors. Conversely, TGS2610 lies near the origin, confirming its negligible contribution to the PCA model.

Overall, the loading structure shows that odor discrimination arises from the combined behavior of sensor groups, with VOC-sensitive sensors dominating intensity (PC1) and selective sensors contributing to compositional differentiation (PC2), supporting the effectiveness of the heterogeneous adaptive e-nose design.

3.2.5. Discussion of Integrated Multivariate Sensor Responses

Bringing together the ANOVA and PCA results provides a clearer picture of how the electronic nose system behaves as a whole rather than as isolated sensors. While the ANOVA analysis shows that individual sensors can distinguish between fruit samples, it does not fully reflect the complexity of odor detection, which is inherently a multivariate process.

The PCA analysis shows how the first two components capture most of the variation in the dataset and present well-defined clusters for the different fruits. Differences in overall VOC intensity are presented along PC1, with strawberry and mango showing stronger responses, while apple and pear show weaker emissions. The consistency between these trends and the ANOVA results gives confidence that the patterns observed are robust.

Looking at sensor behavior more closely, it becomes evident that not all sensors contribute equally. VOC-sensitive sensors, particularly BME688, SGP30, and SGP41, play a central role in defining the main patterns in the data, effectively capturing the intensity of the aroma. At the same time, selective sensors such as NO_x, NO, H₂S, and the MEMS channels provide additional detail by responding to differences in gas composition. The loading plots help to make this relationship clearer, showing how VOC-driven responses are separated from oxidizing gases such as NO_x and O₃.

This also allows a simple but useful ranking of the sensors. VOC-focused sensors emerge as the most influential group, followed by selective sensors that help refine the discrimination between samples. Other sensors play a more limited role, and in some cases, such as TGS2610, contribute very little to the overall model.

3.2.6. Implications for Adaptive e-Nose Design

The multivariate analysis confirms that effective odor discrimination emerges from the collective response of heterogeneous sensors, rather than reliance on any single sensing element. VOC-focused MOS sensors provide sensitivity to overall aroma intensity, while electrochemical sensors targeting NOx, NO, H₂S, and CO contribute selectivity and compositional resolution.

Importantly, the PCA results validate the adaptive e-nose framework proposed in this study: newly integrated sensors can be quantitatively evaluated not only in isolation but also in terms of their multivariate contribution to odor classification. Sensors exhibiting consistently high communalities and strong influence on principal components represent optimal candidates for future deployments, while low-impact sensors may be excluded to reduce cost, power consumption, and system complexity.

Collectively, these results highlight the complementary strengths of the tested sensors:

• BME688 and SGP30: High sensitivity and excellent repeatability, ideal for quality grading, VOC profiling, and automated ripeness evaluation (future study).
• Multichannel MEMS sensor: Broader chemical responsiveness, valuable for complex mixture differentiation or identifying compositional shifts across cultivars.
• SGP41: Stable VOC index output with strong detection in the optimal range, enabling continuous monitoring and consumer-facing freshness systems.

The agreement between sensor trends and published GC–MS findings on strawberry VOC emissions underscores the promise of low-cost electronic nose (e-nose) technologies as scalable, real-time alternatives for agricultural quality assessment.

4. Conclusions

This study demonstrates that effective odor discrimination in low-cost electronic nose systems is achieved not through individual sensor performance alone, but through the coordinated behavior of a heterogeneous sensor array. By combining MOS, MEMS, electrochemical, and NDIR sensing technologies, the proposed system captures a broad representation of fruit-derived gaseous signatures as a proof of concept.

The analysis proves that VOC-sensitive sensors, such as BME688, SGP30, and SGP41, play a key role in detecting overall aroma intensity, while selective sensors, including NO_x, NO, H₂S, and multichannel MEMS elements, provide additional information related to gas composition and chemical balance. In contrast, certain sensors, such as TGS2610, contribute minimally under the tested conditions.

Importantly, these findings support the concept of an adaptive e-nose framework, where sensor relevance is determined based on multivariate contribution rather than isolated response.

This work proves that low-cost, modular sensor arrays can provide interpretable odor fingerprinting. Future work will focus on integrating advanced machine learning models and evaluating newly emerging sensors to further enhance system adaptability and enable deployment in real-world applications such as food quality monitoring, environmental sensing, and smart agriculture.