Advanced Computational Framework for Electronic Nose Development

A Synopsis
In partial fulfillment of the degree of
Doctor of Philosophy

By

Vyom Kumar Gupta

Enrollment No. PMI2017003

Under the guidance of

Dr. Sunny

Department of Electronics and Communication Engineering

Indian Institute of Information Technology Allahabad
March 2025
1 Introduction
Ambient sensing using electronic noses (E-noses) has gained significant attention in recent decades
for detecting and analyzing changes in environmental conditions. These systems are widely used
in various applications, including environmental monitoring, food processing, fitness tracking, non-
invasive breath analysis for medical diagnostics, indoor air quality assessment, agricultural studies,
and industrial safety. E-noses enable chemical analysis of environments using sensors combined with
decision-making algorithms [1].

Figure 1: An Overview of E-nose System Design

Figure 1 provides an overview of an E-nose system. Gas sensors play a crucial role in E-nose tech-
nology, leveraging advancements in surface science [2], [3]. Common sensor materials include metal
oxides, metal sulfides, and conducting polymers, which can detect volatile organic compounds (VOCs)
and harmful gases. These materials are deposited on substrates using techniques like spin coating and
electrospinning. Among various sensor types, metal oxide semiconductor (MOX) gas sensors are
widely used. They function by changing their electrical properties in response to gas exposure. When
gas molecules interact with the metal oxide surface, a chemical reaction occurs, leading to adsorption
and a measurable change in the sensor’s electrical resistance. This resistance change is converted into
an electrical signal, where the signal intensity is proportional to the gas concentration.
E-nose systems integrate gas sensors with computational units to analyze gas concentrations using
mass flow controllers. The sensor response (measured in volts or ohms) is recorded alongside gas con-
centration levels. This labeled data is then processed using signal transformation techniques to extract
meaningful features. Machine learning (ML) models are trained on these transformed representations
to classify or quantify gas compositions accurately.
Various ML techniques, such as Support Vector Machines (SVM), Logistic Regression (LR), Ran-
dom Forests (RF), k-Nearest Neighbors (KNN), and Neural Networks (NNs), including Artificial NNs,

1
Convolutional NNs (CNNs), and Recurrent NNs (RNNs), have been explored for E-nose applica-
tions [4], [5], [6]. Although NNs are effective in E-nose systems [7], [8], they come with several
challenges. Sensor responses are affected by environmental factors such as humidity and temperature,
leading to sensor surface degradation over time—a phenomenon known as sensor drift [9].
In addition, NNs are prone to overfitting, where models become highly specialized to training data
but fail to generalize to unseen scenarios. While hyperparameter tuning can enhance generalization,
it is a time-consuming and computationally expensive process [10]. Furthermore, NN models are
resource-intensive, requiring significant computational infrastructure. To deploy them in resource-
constrained environments, optimization techniques such as quantization and pruning are essential for
reducing memory and computation requirements [11], [12].

2 Literature
The development of electronic nose (E-nose) systems involves four critical areas of research: sen-
sor data preprocessing, sensor drift mitigation, machine learning model optimization, and resource-
efficient deployment. While significant progress has been made in these domains, challenges remain
in achieving robust, adaptable, and computationally efficient E-nose solutions. This section reviews
key advancements in each area and highlights existing research gaps.

2.1 Sensor Data Preprocessing for E-Nose Systems

Preprocessing sensor data is essential for improving gas classification accuracy by transforming raw
signals into meaningful representations. Various signal transformation techniques, including Principal
Component Analysis (PCA), Independent Component Analysis (ICA), and wavelet decomposition,
have been widely used to extract features from gas sensor responses [13], [14], [15]. PCA and ICA
reduce dimensionality and remove noise, improving model performance by isolating dominant sensor
response patterns. Wavelet decomposition, on the other hand, enhances temporal feature extraction by
capturing both high-frequency and low-frequency signal variations. However, despite their advantages,
these methods do not fully compensate for sensor drift, which remains a major challenge in long-term
E-nose deployment.

2.2 Sensor Drift Mitigation: Challenges and Advances

Sensor drift, caused by environmental factors such as humidity, temperature fluctuations, and sensor
aging, degrades classification performance over time. Addressing this issue is crucial for maintaining
the long-term reliability of E-nose systems. Below are the traditional approaches that promisingly
mitigates the drift issue.

• Regularized Neural Network Ensembles: Verma et al. proposed an ensemble-based neural net-
work model to enhance classification accuracy by aggregating multiple classifiers. However, this
method does not inherently handle domain shift, where the distribution of training (source) and
testing (target) data differs [16].

• Domain Adaptation (DA) Methods: Guo et al. introduced a DA-based approach, achieving an
average accuracy of 76.63% [17], while Liu et al. proposed an open-set DA framework with an
accuracy of 77.70% [18]. These models improve generalization but require labeled or unlabeled
target-domain data, making them impractical for real-world applications where such data is often
unavailable [10].

2
2.3 Machine Learning Optimization for Gas Classification
Traditional gas classification models rely on manually tuned neural networks, which require extensive
hyperparameter optimization to balance accuracy and computational efficiency. While methods such
as Support Vector Machines (SVM), k-Nearest Neighbors (KNN), and Random Forests (RF) have
been used [4], [5], [6], deep learning-based solutions, including Artificial Neural Networks (ANNs),
Convolutional Neural Networks (CNNs), and Recurrent Neural Networks (RNNs), have demonstrated
superior performance [7], [8]. However, manual tuning of deep learning models is time-consuming and
computationally expensive. In order to improve the performance generalization of the neural network
models, research have implemented neural networks and tune the model’s hyper-parameters manually
which is a tedious process.

2.4 Resource-Constrained Deployment of E-Nose Systems

Deploying deep learning models in embedded and real-time environments requires optimizing mem-
ory and computational resources. Existing models demand high computational power, making them
unsuitable for resource-constrained hardware such as Field-Programmable Gate Arrays (FPGA) and
Application-Specific Integrated Circuits (ASICs). To address this, researchers have explored:

• Quantization-Aware Training (QAT): Reduces model precision (e.g., FP32 → INT8 or INT4)
while maintaining accuracy.

• Post-Training Quantization (PTQ): Applies quantization after model training, offering 18-20×
model compression but with potential accuracy loss.

• Network Pruning: Removes redundant connections in neural networks, achieving 4.5-12.3×

compression at the cost of reduced accuracy [11], [12].

While these techniques improve efficiency, there is still a lack of research on optimizing E-nose models
for FPGA and ASIC deployment while maintaining classification accuracy.

3 Research Gaps
Despite advancements in E-nose systems, several key challenges remain:

• Exploration of advanced signal preprocessing techniques for sensor signal transformation.

• Sensor Drift Mitigation Without Target-Domain Data: Existing drift correction models rely on
labeled target-domain data, which is often unavailable in real-world applications. A target-
domain-free adaptation framework is needed.

• Automated Model Optimization: Current neural network models require extensive manual tun-
ing. Integrating Bayesian Optimization-based NAS can streamline architecture selection and
improve performance.

• Efficient On-Chip Deployment: Most E-nose models are computationally expensive and diffi-
cult to deploy on edge devices. A hardware-aware optimization pipeline for FPGA and ASIC
implementation is essential.

3
4 Research Objective
Based on the literature review and identified research gaps, the following key objectives have been
formulated:

• Development of an E-nose System for Gas Classification and Quantification (P1): I) Inves-
tigate electronic nose (E-nose) technology for classifying and quantifying gases/mixtures using
data from in-house fabricated sensors and publicly available datasets. II) Explore advancements
in signal processing and computational methods for feature extraction and low-computational
analysis.

• Sensor Drift Mitigation via Target-Domain Data-Free Domain Adaptation (P2): I) Address
both long-term and short-term sensor drift in gas sensor array datasets using meta-learning-based
domain adaptation, eliminating the need for target-domain data. II) Develop adaptive models that
generalize effectively across drift environments.

• Automated Neural Network Optimization for E-Nose Applications (P3): I) Design and
implement Bayesian Optimization-based Neural Architecture Search (NAS) to develop well-
generalized models for three distinct E-nose applications. II) Optimize hyperparameters and
network structures automatically to enhance model performance.

• Hardware-Efficient Deployment on FPGA and ASIC (P4): I) Implement the optimized neural
network models on FPGA and ASIC platforms. II) Utilize hardware-aware optimization tech-
niques such as quantization, pruning, and high-level synthesis (HLS) to balance computational
efficiency and accuracy.

5 Research Methodology and Work Plan

Figure 2: Schematic Illustration of the Proposed Research Work

The research methodology is designed to systematically address the four research problems (P1,
P2, P3, and P4) by integrating advanced data acquisition, machine learning, and hardware optimization
techniques. The overall workflow is illustrated in Figure 2 and consists of the following key phases:

4
5.1 Data Acquisition and Preprocessing (P1)
• Time-series gas sensor data is collected from an in-house fabricated metal oxide semiconductor
(MOX) sensor for volatile organic compound (VOC) classification under temperature modula-
tion. Additionally, a publicly available vehicular exhaust dataset is utilized for gas classification
and quantification in a quaternary mixture.

• A peak detection-based feature extraction method is applied to convert time-series sensor re-
sponses into static data representations, ensuring a more structured and meaningful input for
machine learning models.

• To enhance classification and quantification efficiency in vehicular exhaust dataset, a Class-

Distribution-Aware Unified Prediction Model is employed, which first classifies the gas mixture
into distinct categories (Layer-1) and then quantifies the concentration of each gas component
within the identified category (Layer-2). This hierarchical approach reduces computational over-
head by utilizing a minimal number of regression models per gas, improving both efficiency and
accuracy.

5.2 Sensor Drift Mitigation and Adaptation (P2)

• The publicly available gas sensor array drift dataset is analyzed to understand the effects of
long-term and short-term sensor drift on classification performance.

• A Model-Agnostic Meta-Learning (MAML) framework is implemented for target-domain-free

adaptation, allowing the model to generalize without requiring labeled target-domain data.

• The adaptation process enhances robustness against sensor drift, ensuring reliable real-world
deployment.

5.3 Neural Network Optimization with AutoML (P3)

• A Bayesian Optimization-based Neural Architecture Search (NAS) framework is developed to
automate the design of NN models for gas classification in diverse E-nose applications.

• The NAS framework explores a bounded hyperparameter search space and iteratively selects the
best-performing NN architectures based on classification accuracy and computational efficiency.

• The final models are optimized to minimize overfitting and computational complexity while
maintaining high accuracy.

5.4 Hardware Implementation and Resource Optimization (P4)

• The optimized NN models are deployed on FPGA and ASIC platforms to enable real-time gas
classification with minimal computational overhead.

• Hardware-aware optimization techniques such as Quantization-Aware Training (QAT), Post-

Training Quantization (PTQ), and structured pruning are applied to balance the trade-off between
accuracy, latency, and resource constraints. The final hardware implementation is validated for
metrics such as power consumption, inference latency, and computational efficiency, ensuring
compatibility with embedded systems.

The trained and optimized models are tested across three distinct applications—beef quality ex-
amination (food safety), hazardous gas detection (industrial safety), and vehicular exhaust monitoring

5
(environmental assessment)—to evaluate real-world generalization. Performance is assessed using
key metrics such as classification accuracy, F1-score, computational efficiency, and memory footprint
to ensure practical feasibility. This research methodology establishes a comprehensive end-to-end
pipeline that addresses challenges in sensor drift compensation, automated neural architecture search,
and resource-efficient model deployment, enabling robust, real-time gas classification for diverse ap-
plications.

6 Specific Contribution and Novelty

6.1 Peak Detection Assisted Classification of VOCs Using Temperature Modu-
lated SnO2-ZnO Sensor
A novel peak detection-based feature extraction method is developed to enhance classification accuracy
by converting time-series sensor responses into static data representations. The approach utilizes a
composite SnO2 -ZnO sensor, modulated at temperatures ranging from 120°C to 250°C, to capture
distinct gas response patterns. The preprocessed data is classified using supervised learning models,
achieving an accuracy of up to 100%.

6.2 Class-Distribution-Aware Unified Prediction Model for Quantification of

Quaternary Mixture of Vehicular Exhausts
A two-layered unified prediction model is proposed to classify (Layer-1) and quantify (Layer-2) gas
concentrations in a quaternary vehicular exhaust mixture with minimal computational overhead. This
model reduces complexity by utilizing only one regression model per gas, improving efficiency. Exper-
imental evaluation on Google Colab using Python programming demonstrates that K-Nearest Neigh-
bors (KNN) achieves the highest classification accuracy (96% F1-score) with a minimum compu-
tational time of 0.39 seconds. Additionally, eXtreme Gradient Boosting (XGBoost) and Multilayer
Perceptron (MLP) models effectively quantify NO, C3 H8 , NO2 , and NH3 , with coefficient of determi-
nation (R2 ) scores of 0.71, 0.97, 0.95, and 0.94, respectively.

6.3 Target-Domain Data Free Model-Agnostic-Meta-Learning Framework For

Time-Varying Drift Correction For E-nose
A Model-Agnostic Meta-Learning (MAML) framework is developed to mitigate both long-term and
short-term sensor drift without requiring labeled target-domain data. This approach ensures the model
remains adaptable to evolving sensor response distributions, avoiding overfitting to specific scenarios.
Experimental validation on a 36-month dataset with time-varying drift shows an improvement in aver-
age recognition accuracy to 81.64% for long-term drift and 89.62% for short-term drift, significantly
outperforming existing drift correction techniques.

6.4 Bayesian Optimization Based Neural Architecture Search for Classification

of Gases/Odors Mixtures
A Bayesian Optimization (BO)-based Neural Architecture Search (NAS) framework is introduced to
automate the design of optimized NN models for diverse E-nose applications. By modeling the search
space with a Gaussian distribution and minimizing categorical cross-entropy loss, the framework effi-
ciently selects high-performing neural architectures. The optimized models achieve exceptional clas-
sification accuracy across different applications:
• 100% for hazardous gas detection,

6
 • 99.68% for beef quality assessment, and

• 99.46% for vehicular exhaust monitoring.

6.5 On-Chip Implementation of Neural Network Based Classifier Models For

E-nose With Chemometric Analysis
To enable real-time, resource-efficient deployment, the trained NN models are implemented on FPGA
and ASIC platforms using a high-level synthesis (HLS) workflow. The models, initially developed in
Python, are redefined in C++ and synthesized into optimized RTL logic for FPGA/ASIC implementa-
tion. Hardware-aware techniques such as quantization and pruning significantly reduce memory and
computational overhead while maintaining high classification accuracy. The optimized models demon-
strate compact, low-power operation with no preprocessing overhead, making them highly suitable for
edge computing applications.

7 Research Achievements
The research successfully addresses critical challenges in gas classification, quantification, sensor drift
compensation, neural network optimization, and hardware deployment for electronic nose (E-nose)
applications. The key achievements are as follows:

• High-Accuracy Classification of VOCs: The peak detection-assisted preprocessing approach ap-

plied to temperature-modulated SnO2 -ZnO sensor responses effectively transformed time-series
data into structured features, achieving 100% classification accuracy for volatile organic com-
pounds (VOCs).

• Efficient Gas Quantification in Vehicular Exhaust Mixtures: The Class-Distribution-Aware Uni-

fied Prediction Model successfully classified and quantified a quaternary mixture of NO, C3 H8 ,
NO2 , and NH3 . The model’s hierarchical two-layer structure reduced computational overhead,
with K-Nearest Neighbors (KNN) achieving 96% classification accuracy and XGBoost/MLP
models attaining R2 scores of 0.71, 0.97, 0.95, and 0.94, respectively.

• Significant Improvement in Sensor Drift Mitigation: The Model-Agnostic Meta-Learning based

drift correction framework effectively mitigated both long-term and short-term drift without re-
quiring labeled target-domain data. The proposed method outperformed existing drift correction
techniques, improving long-term accuracy to 81.64% and short-term accuracy to 89.62% on a
36-month drift dataset.

• Robust and Optimized Neural Network Architectures: The Bayesian Optimization-based Neural
Architecture Search (NAS) framework successfully automated the design of optimized neural
networks for E-nose applications. The NAS-designed models achieved outstanding classification
accuracies of 100% for hazardous gas detection, 99.68% for beef quality assessment, and 99.46%
for vehicular exhaust monitoring, demonstrating their generalizability across diverse datasets.

• Hardware-Efficient Deployment on FPGA and ASIC: Optimized NN models were successfully

implemented on FPGA and ASIC platforms using Quantization-Aware Training (QAT), Post-
Training Quantization (PTQ), and structured pruning, significantly reducing computational over-
head. QAT achieved a 15-20× model compression with an accuracy drop of only 0.5-2%, while
PTQ and pruning enabled further compression at varying accuracy trade-offs. Lowering model
precision from FP32 to INT4 improved memory efficiency without significantly affecting clas-
sification performance.

7
 • Optimized Latency and Power Efficiency: The FPGA implementation on a ZCU104 board (100
MHz frequency) achieved an average power consumption of 3.6W, with a performance of 114.3
GOPs (Giga Operations Per Second) at an efficiency of 31.75 GOPs/Watt. Hardware optimiza-
tions, including loop unrolling and pipeline techniques, further reduced execution latency by 23
and 40 cycles, respectively.

8 Conclusion
This research may advance electronic nose (e-nose) technology by addressing key challenges in gas
classification, quantification, sensor drift mitigation, and hardware-efficient deployment. The proposed
peak detection-assisted feature extraction and Class-Distribution-Aware Unified Prediction Model en-
able highly accurate and computationally efficient gas detection. The Model-Agnostic Meta-Learning
(MAML) framework effectively mitigates sensor drift without requiring target-domain data, ensur-
ing long-term adaptability. Additionally, the Bayesian Optimization-based Neural Architecture Search
(NAS) automates NN model optimization, enhancing generalization across diverse applications. By
integrating quantization, pruning, and hardware-aware optimization, the optimized models achieve
low-power, real-time deployment on FPGA and ASIC platforms, making e-nose technology feasible
for edge computing applications. This research establishes a scalable, robust, and resource-efficient
e-nose system with applications in food safety, industrial hazard detection, and environmental mon-
itoring. With its strong foundation in machine learning, adaptive drift correction, and hardware effi-
ciency, this work paves the way for the next generation of intelligent, scalable, and real-time e-nose
systems. Future research can explore real-time adaptation, miniaturized hardware implementations,
and expanded datasets to further enhance the robustness and deployment scope of e-nose technology.

9 List of Publications
SCI/SCIE Journal Publications:
1. Vyom Kumar Gupta, S. K. Lalwani, G. S. Bhati, S. Prakash and Sunny, ”Bayesian Optimiza-
tion Based Neural Architecture Search for Classification of Gases/Odors Mixtures,” in IEEE
Sensors Journal,vol.24,no.5,pp.7119-7125,1March,2024,doi: 10.1109/JSEN.2024.3349862 Im-
pact factor = 4.32
2. Vyom Kumar Gupta, A. Yadav, M. Kumar, B. Kumar and Sunny, ”On-Chip Implementation
of Neural Network Based Classifier Models For E-nose With Chemometric Analysis,” in IEEE
Transactions on Instrumentation and Measurement,doi:10.1109/TIM.2025.3554292.Impact
Factor = 5.6
3. Vyom Kumar Gupta, G.S. Bhati, S.K. Lalwani, and Sunny, ”Target-Domain Data Free Model-
Agnostic-Meta-Learning Framework For Time-Varying Drift Correction For E-nose”. in IEEE
Sensor Letters, (2024).[Under Revision] Impact Factor = 2.2
4. Vyom Kumar Gupta, and Sunny, ”Variational-Inference Assisted Designing of Neural Net-
works for Time-varying Drift Correction in E-nose”, in IEEE Sensors Journal, [To be Com-
municated] Impact factor = 4.32
Conference Publications:
1. Vyom Kumar Gupta, S. K. Lalwani, S. Prakash and Sunny, ”Peak Detection Assisted Classi-
fication of VOCs Using Temperature Modulated SnO2 -ZnO Sensor”,IEEE 23rd International
Conference on Nanotechnology (NANO), Jeju City, Korea, Republic of, 2023, pp. 443-446,
doi: 10.1109/NANO58406.2023.10231227.

8
 2. Vyom Kumar Gupta, G. S. Bhati, A. Yadav, V. Yadav, S. Prakash, and Sunny, ”Class-Distribution-
Aware Unified Prediction Model for Quantification of Quaternary Mixture of Vehicular Ex-
hausts”, 4th International Conference on Artificial Intelligence and Signal Processing (AISP),
VIJAYAWADA, India, 2024.

Other Publications:

1. Lalwani, Suraj Kumar, Ajit Debnath, Vyom kumar Gupta, and Sunny. ”On optimization of
electrospun SnO2 -ZnO nanofibers for low concentration ethanol sensing. ”Journal of Materials
Science: Materials in Electronics 34, no. 24 (2023): 1693. Impact Factor = 2.7

2. A. Yadav, Vyom Kumar Gupta and B. Kumar, ”Lightweight Surveillance Image Classifi-
cation Through Hardware-Software Co-Design,” in IEEE Embedded Systems Letters, doi:
10.1109/LES.2025.3534237. Impact Factor = 1.7

References
[1] H. Chen, D. Huo, and J. Zhang, “Gas recognition in e-nose system: A review,” IEEE Transactions
on Biomedical Circuits and Systems, vol. 16, no. 2, pp. 169–184, 2022.

[2] G. Neri, “First fifty years of chemoresistive gas sensors,” Chemosensors, vol. 3, no. 1, pp. 1–20,
2015. [Online]. Available: https://www.mdpi.com/2227-9040/3/1/1

[3] W. Liu, M. Haardt, M. S. Greco, C. F. Mecklenbräuker, and P. Willett, “Twenty-five years of

sensor array and multichannel signal processing: A review of progress to date and potential
research directions,” IEEE Signal Processing Magazine, vol. 40, no. 4, pp. 80–91, 2023.

[4] W. Zhang, L. Wang, J. Chen, X. Bi, C. Chen, J. Zhang, and V. Hans, “A novel gas recognition and
concentration estimation model for an artificial olfactory system with a gas sensor array,” IEEE
Sensors Journal, vol. 21, no. 17, pp. 18 459–18 468, 2021.

[5] L. Zhao, X. Li, J. Wang, P. Yao, and S. A. Akbar, “Detection of formaldehyde in mixed vocs gases
using sensor array with neural networks,” IEEE Sensors Journal, vol. 16, no. 15, pp. 6081–6086,
2016.

[6] R. Gutierrez-Osuna, “Pattern analysis for machine olfaction: a review,” IEEE Sensors Journal,
vol. 2, no. 3, pp. 189–202, 2002.

[7] F. Shaheen, B. Verma, and M. Asafuddoula, “Impact of automatic feature extraction in deep learn-
ing architecture,” in 2016 International Conference on Digital Image Computing: Techniques and
Applications (DICTA), 2016, pp. 1–8.

[8] B. Chikhaoui and F. Gouineau, “Towards automatic feature extraction for activity recognition
from wearable sensors: A deep learning approach,” in 2017 IEEE International Conference on
Data Mining Workshops (ICDMW), 2017, pp. 693–702.

[9] A. B. Suraj Kumar and Sunny, “Baseline drift improvement through investigating a novel ag
doped sno2/zno nanocomposite for selective ethanol detection,” IEEE Transactions on Nanotech-
nology, vol. 18, pp. 412–420, 2019.

[10] H. Se, K. Song, H. Liu, W. Zhang, X. Wang, and J. Liu, “A dual drift compensation framework
based on subspace learning and cross-domain adaptive extreme learning machine for gas sensors,”
Knowledge-Based Systems, vol. 259, p. 110024, 2023.

9
[11] A. Ait Si Ali, H. Djelouat, A. Amira, F. Bensaali, M. Benammar, and
A. Bermak, “Electronic nose system on the zynq soc platform,” Micropro-
cessors and Microsystems, vol. 53, pp. 145–156, 2017. [Online]. Available:
https://www.sciencedirect.com/science/article/pii/S0141933116302691

[12] F. Sun, B. Wu, J. Yan, S. Duan, and X. Hu, “Electronic nose pattern recognition engine: Design,
build, and deployment,” IEEE Sensors Journal, vol. 23, no. 2, pp. 1507–1514, 2023.

[13] X.-Y. Tian, Q. Cai, and Y.-M. Zhang, “Rapid classification of hairtail fish and pork freshness
using an electronic nose based on the pca method,” Sensors, vol. 12, no. 1, pp. 260–277, 2012.
[Online]. Available: https://www.mdpi.com/1424-8220/12/1/260

[14] W. Yan, Z. Tang, G. Wei, J. Yang, L. Province, and S. P. Yantai, “An electronic nose recognition
algorithm based on pca-ica preprocessing and fuzzy neural network,” matrix, vol. 1, p. 2, 2012.

[15] T. Liu, W. Zhang, J. Li, M. Ueland, S. L. Forbes, W. X. Zheng, and S. W. Su, “A multiscale
wavelet kernel regularization-based feature extraction method for electronic nose,” IEEE Trans-
actions on Systems, Man, and Cybernetics: Systems, vol. 52, no. 11, pp. 7078–7089, 2022.

[16] M. Verma, S. Asmita, and K. Shukla, “A regularized ensemble of classifiers for sensor drift
compensation,” IEEE Sensors Journal, vol. 16, no. 5, pp. 1310–1318, 2015.

[17] T. Guo, X. Tan, L. Yang, Z. Liang, B. Zhang, and L. Zhang, “Domain adaptive
subspace transfer model for sensor drift compensation in biologically inspired electronic
nose,” Expert Systems with Applications, vol. 208, p. 118237, 2022. [Online]. Available:
https://www.sciencedirect.com/science/article/pii/S0957417422013847

[18] T. Liu, Y. Wang, and H. Wang, “Open set domain adaptation for electronic nose drift compensa-
tion on uncertain category data,” IEEE Transactions on Instrumentation and Measurement, 2024.

10