electronics

Article
Gas Leakage Detection Using Tiny Machine Learning
Majda El Barkani 1, *, Nabil Benamar 1,2, , Hanae Talei 1 and Miloud Bagaa 3, *

1 School of Science and Engineering, Al Akhawayn University, Hassan II Avenue,

Ifrane 53000, Morocco; [email protected] (N.B.); [email protected] (H.T.)
2 School of Technology, Moulay Ismail University of Meknes, Meknes 50050, Morocco
3 Department of Electrical and Computer Engineering, University of Quebec at Trois-Rivières,
Trois-Rivieres, QC G8Z 4M3, Canada
* Correspondence: [email protected] (M.E.B.); [email protected] (M.B.)

Abstract: Gas leakage detection is a critical concern in both industrial and residential settings, where
real-time systems are essential for quickly identifying potential hazards and preventing dangerous
incidents. Traditional detection systems often rely on centralized data processing, which can lead
to delays and scalability issues. To overcome these limitations, in this study, we present a solution
based on tiny machine learning (TinyML) to process data directly on devices. TinyML has the
potential to execute machine learning algorithms locally, in real time, and using tiny devices, such as
microcontrollers, ensuring faster and more efficient responses to potential dangers. Our approach
combines an MLX90640 thermal camera with two optimized convolutional neural networks (CNNs),
MobileNetV1 and EfficientNet-B0, deployed on the Arduino Nano 33 BLE Sense. The results show
that our system not only provides real-time analytics but does so with high accuracy—88.92% for
MobileNetV1 and 91.73% for EfficientNet-B0—while achieving inference times of 1414 milliseconds
and using just 124.8 KB of memory. Compared to existing solutions, our edge-based system overcomes
common challenges related to latency and scalability, making it a reliable, fast, and efficient option.
This work demonstrates the potential for low-cost, scalable gas detection systems that can be deployed
widely to enhance safety in various environments. By integrating cutting-edge machine learning
models with affordable IoT devices, we aim to make safety more accessible, regardless of financial
limitations, and pave the way for further innovation in environmental monitoring solutions.

Keywords: CNN; cost-effective IoT solutions; EfficientNet-B0; IoT; MobileNetV1; TinyML

Citation: El BarkaniEl Barkani, M.;
Benamar, N.; Talei, H.; Bagaa, M. Gas
Leakage Detection Using Tiny
Machine Learning. Electronics 2024, 13,
1. Introduction
4768. https://doi.org/10.3390/
electronics13234768 The Internet of Things (IoT) has emerged as a transformative technology that funda-
mentally reshapes our interaction with the digital world. By seamlessly integrating sensors
Academic Editor: Marcin Witczak
and actuators into everyday objects, the IoT has enabled a new era of connectivity and
Received: 21 October 2024 real-time data collection in various sectors. From enhancing home automation to revolu-
Revised: 21 November 2024 tionizing industrial operations, the deployment of IoT technologies promises improved
Accepted: 26 November 2024 efficiency and responsiveness in real-time applications [1].
Published: 2 December 2024 The merging of AI with the Internet of Things (IoT) expands the potential of IoT
systems. This connection enables advanced data analytics and decision-making processes
at network edges, eliminating the latency and bandwidth constraints associated with cloud
computing. The AIoT not only improves decision making based on IoT data but also enables
Copyright: © 2024 by the authors.
complex applications like predictive maintenance and smart energy management [2].
Licensee MDPI, Basel, Switzerland.
Despite the numerous benefits, deploying AIoT systems is not without problems. Some of
This article is an open access article
distributed under the terms and
the significant challenges include high implementation costs, complexity in integrating AI
conditions of the Creative Commons
with current IoT infrastructures, and concerns about data privacy and security. Furthermore,
Attribution (CC BY) license (https://
the continual functioning of these systems requires sustainable energy, which can be a
creativecommons.org/licenses/by/ limiting factor, especially in environments where power availability is constrained [1,2].
4.0/).

Electronics 2024, 13, 4768. https://doi.org/10.3390/electronics13234768 https://www.mdpi.com/journal/electronics

Electronics 2024, 13, 4768 2 of 15

Tiny machine learning (TinyML) has emerged as a potent solution to these challenges.
It represents a fast-growing field of machine learning technologies and applications that
are capable of performing on-device analytics at extremely low power. This capability is
crucial for extending the battery life of IoT devices and reducing latency in data processing.
TinyML enables the embedding of intelligence in the tiniest devices, making it feasible to
process data directly on the device rather than relying on cloud services [2,3].
Recent studies further underscore TinyML’s transformative impact on various appli-
cations. For instance, Abadade et al. [4] provided a comprehensive survey of TinyML,
highlighting its advantages of energy efficiency, low latency, and privacy. The survey dis-
cussed TinyML’s suitability for resource-constrained devices in sectors such as healthcare,
environmental monitoring, and anomaly detection, showing how it brings machine learn-
ing capabilities to edge devices in diverse fields. Another study by Lin et al. [5] explored
the challenges associated with deploying deep learning models on microcontrollers due to
memory limitations. They emphasized the need for system-algorithm co-design to make
TinyML feasible for real-time applications on ultra-low-power devices, enabling impactful
IoT deployments across various industries.
In addition, recent research by Pimpalkar and Niture [6] presented a TinyML-based
contactless elevator system that utilizes convolutional neural networks (CNNs) for person
detection and keyword spotting. This innovative approach highlights TinyML’s role in
public health and safety applications by reducing physical contact, which is essential in
contexts where hygiene and convenience are paramount. The study serves as a practical
example of how TinyML can be implemented in real-world systems with low computational
overhead and minimal infrastructural changes.
TinyML provides an efficient framework for deploying machine learning models on
cost-effective and eco-friendly platforms such as the Arduino Nano 33 BLE. By leveraging
techniques such as model quantization and on-device learning, TinyML ensures that
real-time analytics can be performed while minimizing power consumption, making it
a crucial technology for safety applications like gas leak detection. These techniques,
including quantization and pruning, are integral in reducing model size and computational
requirements, allowing deployment on resource-constrained devices without sacrificing
accuracy [7,8].
Specifically, in the realm of safety and environmental monitoring, TinyML can be
instrumental [9]. Gas leaks, particularly in industrial settings or residential areas, pose sig-
nificant risks, including explosions and health hazards due to toxic exposures. Traditional
gas detection systems often rely on central processing and are not capable of real-time
responses. Utilizing TinyML, we propose a novel approach to gas leak detection, where
IoT devices can instantly analyze environmental data and provide alerts without needing
to communicate with a central server. For instance, research in acoustic monitoring has
shown that embedding CNN-based TinyML models into edge devices can achieve high
accuracy and low latency for detecting critical anomalies in sensor data [7]. This method
enhances the speed and reliability of gas leak detection, which is crucial for safety and
preventive measures [1].
To address the critical need for efficient, real-time gas leak detection in resource-
constrained environments, this research aims to design and implement a TinyML-based
gas detection system. The primary goal is to leverage low-power, edge-based machine
learning models to achieve high accuracy and quick response times in detecting various
gas types. Specifically, we focus on integrating the MobileNetV1 and EfficientNet-B0
models with the Arduino Nano 33 BLE Sense and the MLX90640 thermal camera to create
a scalable, cost-effective solution for industrial and residential applications. This work
not only demonstrates the feasibility of deploying TinyML models for safety monitoring
but also contributes to advancing IoT-based environmental monitoring technologies by
providing a low-cost, reliable solution for gas leak detection.
This study addresses three key research questions (RQs) on leveraging TinyML for
efficient, real-time gas leak detection:
Electronics 2024, 13, 4768 3 of 15

• RQ1: How effectively can TinyML models improve detection in resource-limited

environments?
• RQ2: What balance can be achieved between accuracy, speed, and efficiency on edge
devices?
• RQ3: How well do optimized models, such as MobileNetV1 and EfficientNet-B0,
perform under various conditions and hardware setups?
These questions aim to bridge the gap between TinyML’s theoretical advancements
and practical, cost-effective safety solutions.
The remainder of this paper is organized as follows. Section 2 reviews related works,
highlighting advances in the applications of IoT and TinyML for environmental safety.
Section 3 describes the materials and methods, detailing the experimental setup for TinyML-
based gas leak detection. Section 4 presents our findings, discussing the performance of the
proposed system under various conditions. Finally, Section 5 concludes this paper with a
summary of the obtained results and future research directions.

2. Related Works
Numerous studies have explored using machine learning models and TinyML tech-
nologies to enhance gas leak detection in various settings, showcasing advances in identify-
ing leaks with precision and improving safety measures effectively.
Kopbayeva et al. [10] introduced an innovative method to detect and diagnose natural
gas leaks, employing a combination of convolutional networks and bidirectional long- and
short-term memory (BiLSTM) networks. The model impressively achieved a 92% accuracy
rate in predicting leaks and accurately classifying their size, with just a 5% average error
margin. However, the reliance on simulated data poses limitations and could overlook
real-world intricacies. Furthermore, while the model effectively detected leaks within a
50-s time frame, its accuracy waned with shorter intervals, dropping below optimal levels
when the window was reduced to less than 26 s.
Miao et al. [11] used unsupervised learning with the Wasserstein generative adversarial
network with gradient penalty (WGAN-GP) and the Bayesian Gaussian mixture model
(BGMM) to analyze real-time stress signals obtained from residual magnetic effects. WGAN-
GP extracts features from normal stress data, which are then inputted into the BGMM to
assess pipeline health via the WLP indicator. The results indicated that the relative error of
the stress prediction model was within 3%, and the weighted logarithm probability (WLP)
effectively distinguished between normal and leakage conditions. However, limitations
include the need for further evaluation under high pressure and external interference
conditions and consideration of additional potential leakage causes beyond those addressed
in the model.
In their study, Doshmanziari et al. [12] introduced a model-based fault detection
framework for gas pipeline leak detection, employing sensor fusion and a state-space
representation approach. Using the extended Kalman Filter (EKF) for state observation,
the methodology requires hydraulic pressure measurements at various points and recom-
mends using sensor arrays over single sensors for improved leak estimation accuracy. When
validated through simulations using the OLGA software, the approach effectively identified
and located leaks, demonstrating notable precision. However, the performance is sensitive
to factors such as spatial discretization, initial conditions, leak timing, and sampling rate,
indicating areas for potential improvement.
Du et al. [13] developed a deep learning framework for detecting underground natural
gas micro-leaks using hyperspectral imagery, grounded in a multi-branch deep learning
network (Img-Spec-PGE model) that estimates the plant leaf area index (LAI) under gas stress.
Despite achieving high LAI estimation accuracy and validating the method’s efficacy through
realistic simulation models, the study faced notable limitations. The approach’s performance
is particularly sensitive to the selection of spectral bands, implying that its applicability may
be constrained across different vegetation types or environmental conditions.
Electronics 2024, 13, 4768 4 of 15

In the study by Gkogkidis et al. [9], a novel application of TinyML technology was
showcased through the development of a system designed for the real-time detection
of hazardous gas leaks, utilizing edge computing for data processing directly on the
device. This approach circumvents traditional privacy and data transmission concerns
associated with IoT devices by enabling local analysis of sensor data to identify gas leaks
or smoke presence. Despite achieving promising detection outcomes, with F1 scores of
0.77 for smoke and 0.70 for ammonia, the study acknowledged significant limitations,
particularly the challenges associated with offline training or retraining of the system,
security vulnerabilities in data transmission, and the potential need for more specialized
sensors to enhance detection accuracy.
In their innovative research, Lorthong et al. [2] employed an Artificial Neural Network
(ANN) within an IoT-based system to predict the risk of LPG gas leaks, harnessing environ-
mental data captured by sensors. This ANN model, crucial for processing variables like
LPG concentration, temperature, humidity, and oxygen levels, effectively identified com-
plex patterns and relationships to classify leakage risk. Achieving an impressive accuracy
rate of 96.05%, the model underscores the potential of machine learning in enhancing safety
mechanisms in industrial settings. However, the research highlighted limitations, notably
the dependency on high-quality, real-time sensor data and the challenges of applying the
model across diverse and dynamic industrial environments.
In the study by Narkhede et al. [14], a multimodal AI-based sensor fusion approach
was employed for gas detection, integrating a seven-semiconductor gas sensor array with a
thermal camera. This early fusion technique achieved a testing accuracy of 96%, signifi-
cantly outperforming individual sensor models (82% for gas sensors and 93% for thermal
images). However, the method’s reliance on both gas sensors and thermal imaging intro-
duced complexity and increased costs, making it less practical for widespread deployment.
The necessity of using both types of sensors complicates the system, and it can be pro-
hibitively expensive, highlighting the need for more streamlined and cost-effective solutions
in gas detection technology.
Wang et al. [15] introduced the RGB-Thermal Cross-Attention Network (RT-CAN)
for gas detection, combining RGB and thermal imaging to enhance detection accuracy
by 4.86%. They also developed Gas-DB, a dataset containing 1300 RGB-thermal images
for training and evaluation. However, the dual-camera requirement increases costs and
complexity, limiting the model’s use in constrained settings. Additionally, the model’s
reliance on RGB data could impact performance in low-light conditions, and Gas-DB’s
limited environmental diversity may reduce generalizability.
Attallah and Elhelw [16] proposed a gas leak detection pipeline using multiple CNNs
and thermal imaging, achieving 98% accuracy. This method enhances detection but requires
high computational power, making it less suitable for real-time or low-power devices.
The reliance on thermal imaging alone may reduce accuracy in complex environments
where heat interference is present, limiting its adaptability across various gas types.
Sharma et al. [17] applied a multimodal approach and federated learning for gas de-
tection, achieving a 96% accuracy rate while maintaining data privacy. However, federated
learning depends on reliable network connectivity, which could limit deployment in remote
locations. The complexity of processing multimodal data and the specialized setup for
privacy-preserving frameworks may also impact scalability.
Tsoukas et al. [18] developed a TinyML-based gas detection device for real-time
monitoring, achieving F1 scores of 0.77 for smoke and 0.70 for ammonia. While effective,
the offline setup restricts model updates, and limited accuracy may reduce sensitivity
in critical applications. Additionally, the system’s focus on specific gases (smoke and
ammonia) limits broader industrial use.
Shafin et al. [19] introduced a custom CNN-BiLSTM model specifically designed
for gas leak detection on resource-limited IoT devices. The model, developed without
using pre-trained networks, achieved a compact size of 575 KB and an efficient inference
time of 0.255 ms. The combination of CNN layers for feature extraction and BiLSTM
Electronics 2024, 13, 4768 5 of 15

layers for temporal pattern recognition allowed the model to detect gas leaks accurately
while maintaining low latency, making it suitable for real-time safety applications in edge
environments where power and computational resources are limited. Both pre-trained and
custom-designed models can be beneficial for such applications, with the choice depending
on the specific deployment constraints and performance requirements.
In conclusion, the related works highlight the diversity of approaches in gas leak
detection, as detailed in Table 1, which summarizes the models’ accuracies and notes
significant limitations such as dependency on simulated data, high computational demands,
and sensitivity to environmental conditions. This comparison emphasizes the necessity
of addressing these constraints to enhance the practicality and effectiveness of gas leak
detection technologies in real-world applications.

Table 1. Comparison of gas leak detection models.

Reference Model/Method Accuracy Notes

Effective for predicting leaks
Kopbayeva et al. (2022) [10] CNN & BiLSTM 92% and classifying size,
but reliant on simulated data.
Suitable for stress signal
analysis; needs further
Miao et al. (2022) [11] WGAN-GP BGMM Within 3% error
validation under
high-pressure conditions.
Hyperspectral imagery
sensitive to spectral band
Multi-branch DL
Du et al. (2024) [13] High for LAI estimation selection; applicable to
(Img-Spec-PGE)
vegetation types under
gas stress.
F1 scores of Challenges in offline
Gkogkidis et al. (2022) [9] TinyML Edge Detection 0.77 (smoke), retraining and
0.70 (ammonia) security vulnerabilities.
Dependence on high-quality,
real-time sensor data limits
Lorthong et al. (2023) [2] ANN-based IoT System 96.05%
scalability
across environments.
Combines gas sensors and
Multimodal AI-based
Narkhede et al. (2021) [14] 96% thermal imaging, which
sensor fusion
increases cost and complexity.
Requires dual cameras;
RGB-Thermal Cross-
Wang et al. (2024) [15] 4.86% Improvement reliance on RGB may impact
Attention Network
low-light performance.
Compact model suitable for
Shafin et al. (2023) [19] Custom CNN and BiLSTM - IoT deployment without
pre-trained weights.
High computational demand;
Multiple CNNs + Thermal
Attallah and Elhelw (2023) [16] 98% challenging for
Imaging
low-power devices.
Maintains data privacy;
Multimodal Federated
Sharma et al. (2024) [17] 96% depends on
Learning
network connectivity.
Offline setup limits model
TinyML-based F1 score: 0.77 (smoke),
Tsoukas et al. (2023) [18] updates; focused on
Monitoring Device 0.70 (ammonia)
specific gases.
Electronics 2024, 13, 4768 6 of 15

3. Materials and Methods

The following section outlines the materials and methods employed in the study, focus-
ing on data acquisition, preprocessing, and the design of our gas leak detection CNN model
using the Edge Impulse platform. Our approach leverages advanced machine learning
techniques and state-of-the-art hardware to develop an efficient, accurate, and deployable
gas leak detection system.

3.1. Data Acquisition and Preprocessing

In this work, we used an existing dataset curated by Narkhede et al. [14] known as
the “MultimodalGasData” dataset. This dataset is a comprehensive collection designed
specifically for gas detection and classification tasks. It consists of both thermal image
data, which capture temperature variations, and numerical values from gas sensors. This
multimodal approach, combining thermal images with sensor measurements, enhanced the
dataset’s richness and utility, offering a more comprehensive view of the gas leak scenarios.
The experimental setup for data collection was meticulously designed to ensure precise
and consistent data acquisition. The setup involved positioning seven metal-oxide gas
sensors (MQ2, MQ3, MQ5, MQ6, MQ7, MQ8, and MQ135) alongside a Seek Compact
Thermal Imaging Camera (UW-AAA), as shown in Figure 1. These sensors were selected
due to their sensitivity to a range of gases, as shown in Table 2. The sensors were arranged
with a 1 mm separation between each to ensure uniform exposure to the gas sources,
thereby maintaining consistency across the dataset.

Table 2. Gas sensors and sensitive gases.

Sensor Sensitive Gas

MQ2 Methane, Butane, LPG, Smoke
MQ3 Alcohol, Ethanol, Smoke
MQ5 Natural Gas, LPG
MQ6 LPG, Butane Gas
MQ7 Carbon Monoxide
MQ8 Hydrogen Gas
MQ135 Air Quality (Benzene, Smoke)

To simulate realistic gas leak scenarios, perfume spray and incense sticks were chosen
as gas sources. The perfume spray, containing 95% alcohol, and incense sticks, emitting
a mixture of carbon monoxide, carbon dioxide, nitrogen dioxide, and sulfur dioxide,
served as proxies for common indoor pollutants. The controlled release of these gases
provided a dynamic environment, emulating potential leak situations found in industrial
and residential settings.
The data-logging process was structured to capture comprehensive and high-resolution
data. Measurements from the gas sensors and thermal images were recorded at 2 s intervals
over a 90-minute period. This frequent logging ensured that transient changes in the gas
concentration and temperature were accurately captured, providing detailed temporal data
essential for training robust machine learning models.
The release of gases was carefully regulated to ensure consistency. During the first
30 min, gases were introduced every 15 s, then changed to 30 s for the subsequent 30 min,
and then the interval was increased to 45 s in the last 30 min. This systematic approach en-
sured the dataset encompassed a wide range of gas concentrations and thermal signatures,
simulating both rapid and gradual environmental changes.
Electronics 2024, 13, 4768 7 of 15

Figure 1. Experimental setup for data collection (reprinted, with permission, from [20] @2022 MDPI).

The Seek Compact Thermal Imaging Camera was integral to the data acquisition
process, capturing thermal images that reflect temperature variations caused by gas leaks.
The camera, with a resolution of 206 × 156 pixels and a temperature measurement range
from −40 to 330 ◦ C, is capable of detecting thermal anomalies associated with gas emissions.
These images provided crucial visual data that could be processed to identify gas leaks in
visually obscured or inaccessible environments.
The resulting dataset, “MultimodalGasData”, comprises 6400 samples divided equally
into four classes: smoke, perfume, a mixture of smoke and perfume, and a neutral environ-
ment, as shown in Figure 2. Each sample includes both thermal images and corresponding
gas sensor measurements, ensuring a robust and diverse dataset for training and validating
gas leak detection models.

(a) (b) (c) (d)

Figure 2. The four dataset categories: (a) Mixture; (b); No Gas, (c); Perfume; (d) Smoke.

For the purpose of this article, we only considered the thermal image data taken
from the “MultimodalGasData” dataset. This choice was prompted by our desire to avoid
the requirement for sensors in conjunction with the thermal camera. Thermal imaging
has a number of benefits, including the capacity to identify gases in visually obscured or
inaccessible areas, which is critical for early warning systems in residential and industrial
applications. To optimize for edge device deployment, each image was downscaled to 96 ×
96 pixels, preserving key thermal details while reducing computational load. By focusing
our efforts on thermal imaging data, we simplified the complexity of the sensor array and
lowered the cost associated with the deployment of multiple sensor systems, making the
technology available to a larger spectrum of users and applications.

3.2. Gas Leak Detection CNN Model

In this work, we implemented the convolutional neural network (CNN) model for
gas leak detection using the Edge Impulse platform (https://edgeimpulse.com/, accessed
Electronics 2024, 13, 4768 8 of 15

on 28 November 2024), a machine learning model creation, training, and deployment

platform for edge devices. Edge Impulse was selected for this project due to its robustness
in deploying machine learning models on edge devices. The platform offers extensive
capabilities for training, optimizing, designing, and preparing data. For our gas leak
detection system, its transfer learning support and user-friendly interface made it the
perfect option. The flexibility and scalability of the system can be ensured by effectively
preprocessing data, optimizing models, and deploying them across a range of hardware
platforms by using Edge Impulse.
Given that our dataset was in image format, we chose two models—MobileNetV1
and EfficientNet-B0—both of which are deployable on TinyML devices. MobileNetV1 [21]
was selected for its lightweight architecture, allowing it to efficiently process images with
minimal computational power, making it ideal for constrained environments. EfficientNet-
B0 [22], while slightly more resource-intensive, offers higher accuracy and remains efficient
enough for TinyML deployment. By utilizing both models, we ensured that our system
could deliver accurate gas leak detection on a wide range of TinyML devices, from low-
power devices to more capable hardware.
To ensure the models could be deployed efficiently on edge devices, we applied
post-training INT8 quantization. Post-training quantization converts the model’s weights
and activations from 32-bit floating-point to 8-bit integers (INT8), significantly reducing
the memory footprint and computation required for inference. This step was crucial for
enabling the models to run within the memory constraints of the Arduino Nano 33 BLE
Sense, which has only 256 KB of RAM.
Using the Edge Impulse platform’s built-in quantization tool, we optimized the Mo-
bileNetV1 and EfficientNet-B0 models for embedded deployment. The quantization process
reduced the models’ sizes while ensuring that their performance remained close to that of
the original floating-point models. Both models maintained high accuracy after quantiza-
tion, making them viable for real-time gas leak detection on resource-constrained hardware.
For MobileNetV1, we used a version of the system with an input size of 96 × 96
pixels and a depth multiplier of 0.25 for our gas leak detection system. The selection of a
96 × 96-pixel input size for our CNN model was primarily driven by the balance between
computational efficiency and model accuracy. This input size is small enough to meet the
memory constraints of the Arduino Nano 33 BLE Sense, yet large enough to retain essential
spatial information from the thermal images for effective gas detection.
The model was configured with specific settings tailored to our application, as shown
in Table 3.

Table 3. Training configuration and settings.

Parameter Value
Input Axes Image
Image Dimensions 96 × 96 pixels
Color Depth RGB
Training Settings
Number of Training Cycles 45
Learning Rate 0.0005
Training Processor CPU
Validation Set Size 20%
Batch Size 32
INT8 Model Profiling Enabled

The neural network architecture started with an input layer consisting of 27,648 fea-
tures. Each 96 × 96 RGB image resulted in an input layer consisting of 27,648 features
Electronics 2024, 13, 4768 9 of 15

(96 × 96 × 3 channels), providing a sufficient level of detail for the CNN to learn distinctive
features associated with different gas leak scenarios. The pre-trained MobileNetV1 was
fine-tuned to our specific task. The optimization algorithm used was Adaptive Moment
Estimation (Adam), known for its adaptive learning rate capabilities and efficient handling
of sparse gradients. This optimizer combines the advantages of two crucial techniques:
adaptive gradient estimation and momentum, which together improve convergence and
stability during training.
In addition to the primary settings, we employed several key techniques to enhance
model performance and generalization. We used dropout regularization with a rate of 0.5
in the fully connected layers to prevent overfitting. The ReLU activation function was used
to introduce non-linearity into the model, and batch normalization was applied to stabilize
and accelerate the training process.
The second model, EfficientNet-B0, is a state-of-the-art model known for its balance
between model size, computational cost, and performance. However, the configuration
differed slightly from the first model, as we were limited to a maximum of 10 training
cycles. The adopted configuration is presented in Table 4.

Table 4. Training configuration and settings.

Parameter Value
Input Axes Image
Image Dimensions 96 × 96 pixels
Color Depth RGB
Training Settings
Number of Training Cycles 10
Learning Rate 0.0005
Training Processor CPU
Validation Set Size 20%
Batch Size 32
INT8 Model Profiling Enabled

The neural network architecture for EfficientNet-B0 started with an input layer of
27,648 features, and we utilized the pre-trained EfficientNet-B0 without the top layers.
Similar to MobileNetV1, the Adam optimizer was used, leveraging its ability to effectively
handle large-scale data and complex neural networks.
EfficientNet-B0 [22] incorporates several advanced features to enhance performance.
The model employs Swish activation functions, which have been shown to improve ac-
curacy compared to ReLU. Batch normalization is applied throughout the network to
improve training speed and stability. The compound scaling technique, which balances
depth, width, and resolution, ensures that the model is both efficient and effective for a
wide range of applications.
Both MobileNetV1 and EfficientNet-B0 models were implemented using the Edge
Impulse platform, leveraging transfer learning with pre-trained weights optimized for
image recognition tasks. Transfer learning enables these models to retain critical feature
extraction capabilities developed from large, general-purpose datasets, while fine-tuning
adapts the parameters to detect the specific thermal patterns associated with gas leaks in
our ’MultimodalGasData’ dataset. During the fine-tuning process, the models’ final layers
were retrained to capture unique thermal image characteristics linked to gas emissions,
such as subtle variations in heat signatures caused by different gas types. By using pre-
trained architectures with depth multipliers (MobileNetV1 with a 0.25 depth multiplier and
EfficientNet-B0’s compound scaling) and subsequently fine-tuning, we achieved a model
configuration that optimally balances accuracy with computational efficiency.
Electronics 2024, 13, 4768 10 of 15

During training, we monitored key metrics such as loss and accuracy on the validation
set to ensure that the models were learning effectively. We employed early stopping to
prevent overfitting, halting the training process if the validation accuracy did not improve
for 10 consecutive epochs.
By leveraging the MobileNetV1 and EfficientNet-B0 models on the Edge Impulse
platform, we developed a robust and efficient gas leak detection system. MobileNetV1
is ideal for deployment on devices with limited resources due to its lightweight nature,
while EfficientNet-B0 provides higher accuracy suitable for more capable hardware. This
approach ensures that our gas leak detection system can be effectively deployed across
a range of environments, from low-power edge devices to high-performance computing
platforms. The combination of these models allows for versatile and scalable deployment,
ensuring comprehensive coverage and reliability in detecting gas leaks in various settings.
The use of advanced techniques, such as dropout regularization, batch normalization,
and data augmentation, further enhances the performance and generalization of the models,
making them well suited for real-world applications.

3.3. Target Deployment

The goal of this project is to deploy the gas leak detection system on an Arduino
Nano 33 BLE Sense (Cortex-M4F 64 MHz, Arduino, Somerville, MA, USA), complemented
by the MLX90640 thermal camera. This combination is particularly suitable due to the
Arduino’s compact size, low power consumption, and sufficient processing capabilities
for edge machine learning tasks, alongside the MLX90640’s ability to provide detailed
thermal imaging.
The Arduino Nano 33 BLE Sense features an ARM Cortex-M4F processor running at
64 MHz, providing the computational power needed for our CNN models. Key characteris-
tics include 256 KB of RAM and 1 MB of ROM, which are adequate for storing and running
lightweight models like MobileNetV1 and EfficientNet-B0. The MLX90640 thermal camera
offers a 32 × 24-pixel resolution and can measure temperatures ranging from −40 ◦ C to
300 ◦ C, making it ideal for detecting temperature variations associated with gas leaks.
The deployment of the gas leak detection model to the Arduino Nano 33 BLE Sense is
facilitated by the Edge Impulse platform. The deployment process involves several steps,
which are described below.
Model Training and Optimization: The CNN models (MobileNetV1 and EfficientNet-
B0) are trained on the Edge Impulse platform. Transfer learning and post-training quan-
tization techniques are used to ensure that the models are lightweight and efficient for
deployment on resource-constrained devices.
Model Export: After training and validation, the models are exported as a fully
optimized C++ library, which can be directly integrated into an Arduino sketch using the
Edge Impulse SDK. This package includes all necessary code for real-time data acquisition,
preprocessing, and running inference on the Arduino Nano 33 BLE Sense.
Integration with Arduino IDE: The exported model library is integrated into an Ar-
duino sketch using the Arduino Integrated Development Environment (IDE). This inte-
gration involves including the model’s header and source files in the project. The Edge
Impulse SDK is used to facilitate data acquisition, preprocessing, and inference, which are
also incorporated into the sketch.
Code Implementation: The Arduino sketch initializes the onboard camera and collects
real-time data from the MLX90640 thermal camera. The collected data are preprocessed to
match the input format required by the CNN models. The preprocessed data are then fed
into the CNN model to perform inference and detect the presence of gas leaks. Based on the
model’s output, the system can trigger alarms, send notifications via the BLE, or activate
connected devices to mitigate the gas leak.
Optimization Techniques: Using the Edge Impulse platform, we apply post-training
quantization to convert the model’s 32-bit floating-point parameters into 8-bit integers
(INT8). This reduces the model’s memory footprint, making it feasible for deployment on
Electronics 2024, 13, 4768 11 of 15

the Arduino Nano 33 BLE Sense while also speeding up inference. The quantized models
require less memory and computation, resulting in faster predictions and reduced power
consumption, which are critical for real-time applications on a microcontroller.
Testing and Validation: The integrated system is thoroughly tested to ensure accurate
and reliable performance in real-world scenarios. This involves validating the model’s
predictions and the system’s responsiveness under various environmental conditions and
gas concentrations.
Deployment: Once validated, the system is deployed in the target environment.
The compact size and low power requirements of the Arduino Nano 33 BLE Sense, along
with the MLX90640 thermal camera, allow for easy installation in residential, industrial,
or commercial settings where gas leak detection is critical.
By leveraging the Edge Impulse platform, the Arduino Nano 33 BLE Sense, and the
MLX90640 thermal camera, we developed a robust, efficient, and scalable gas leak detection
system. This deployment strategy ensures that our system can be effectively utilized across
various environments, providing enhanced safety and early warning capabilities through
the real-time monitoring and detection of gas leaks.

4. Results
To train and validate the CNN models, we employed a data split of 70% for training,
20% for validation, and 10% for testing. This approach ensured that the models received
ample and diverse training data while providing a robust assessment of their generaliza-
tion performance. Both the MobileNetV1 96 × 96 0.25 and EfficientNet-B0 models were
evaluated through this process, and their performance metrics are detailed below.
The MobileNetV1 model demonstrated an accuracy of 90.2% on the validation set
during the training phase, along with a matching loss value of 0.27. The validation set’s
confusion matrix shows how well the model performed in classifying the various kinds
of gases. In particular, 99.6% of the Mixture class, 83.9% of the No Gas class, 88.4% of the
Perfume class, and 89.1% of the Smoke class were accurately categorized by the model.
For these classes, the F1 scores were 1.00, 0.84, 0.86, and 0.91, in that order.
MobileNetV1’s efficiency was demonstrated by its on-device performance measurements,
which showed an inference time of 1414 ms, a peak RAM utilization of 124.8 KB, and a flash
usage of 304.1 KB. These measurements are essential for ensuring that the model can function
well with the Arduino Nano 33 BLE Sense’s constrained computing capabilities.
During the testing stage, the accuracy of the MobileNetV1 model was 88.92%. The model
successfully categorized 99.4% of the Mixture class, 81.5% of the No Gas class, 84.5% of
the Perfume class, and 90.2% of the Smoke class, according to the confusion matrix shown
in Figure 3. The F1 scores were 1.00, 0.86, 0.88, and 0.92, in that order, showing good
performance in every class, even if the accuracy was somewhat lower than during the
training phase.
With a loss value of 0.28, EfficientNet-B0 demonstrated excellent training accuracy,
attaining 89.5% on the validation set. EfficientNet-B0 accurately categorized 99.6% of the
Mixture class, 83.5% of the No Gas class, 86.2% of the Perfume class, and 88.7% of the
Smoke class, according to the confusion matrix. For these classes, the F1 scores were 1.00,
0.83, 0.85, and 0.91, in that order.
The on-device performance metrics for EfficientNet-B0 were similar to those of Mo-
bileNetV1, with an inference time of 1414 ms, a peak RAM usage of 124.8 KB, and a
flash usage of 304.1 KB. These results confirm that EfficientNet-B0 is also well suited for
deployment on resource-constrained devices.
During the testing phase, the EfficientNet-B0 model achieved a higher accuracy of
91.73%. The confusion matrix for the testing set showed that the model correctly classified
99.7% of the Mixture class, 81.5% of the No Gas class, 90.6% of the Perfume class, and 95.3%
of the Smoke class. The F1 scores were 1.00, 0.88, 0.91, and 0.96, respectively, as shown
in Figure 4, demonstrating the model’s superior performance during testing compared
to MobileNetV1. The confusion matrices utilize a color coding to visually delineate different
Electronics 2024, 13, 4768 12 of 15

accuracy levels: green for the highest percentages of correct predictions, pink for significant
accuracies, light pink for moderate accuracies, and very light pink for lower accuracies and
misclassifications.
Both MobileNetV1 96 × 96 0.25 and EfficientNet-B0 exhibited strong performance in
gas leak detection. However, EfficientNet-B0 marginally outperformed MobileNetV1 in
the testing phase, achieving an accuracy of 91.73% compared to MobileNetV1’s 88.92%.
EfficientNet-B0 also demonstrated higher F1 scores for most classes during testing, indicat-
ing better precision and recall, as shown in Table 5.
The resource usage and efficiency of both models were optimized using post-training
quantization, which reduced their memory footprints and improved their inference speeds,
as demonstrated in Table 6. This optimization is critical for deployment on the Arduino
Nano 33 BLE Sense, which has limited computational resources. Despite their different
architectures, both models demonstrated similar on-device performance metrics, making
them suitable for deployment on the target hardware.

Figure 3. Confusion matrix for the MobileNetV1 model.

Figure 4. Confusion matrix for the BO configuration.

Table 5. Performance comparison of the MobileNetV1 and EfficientNet-B0 models during the
validation and testing phases.

Model Accuracy Precision Recall F1 Score

MobileNetV1
Validation 90.2% 91.8% 89.3% 90.2%
Testing 88.92% 90.6% 87.8% 88.9%
EfficientNet-B0
Validation 89.5% 92.1% 90.0% 90.5%
Testing 91.73% 92.6% 90.7% 91.3%

Table 6. Resource usage metrics for the MobileNetV1 and EfficientNet-B0 models on the Arduino
Nano 33 BLE Sense.

Model Inference Time (ms) Peak RAM Usage (KB) Flash Usage (KB)
MobileNetV1 1414 124.8 304.1
EfficientNet-B0 1414 124.8 304.1
Electronics 2024, 13, 4768 13 of 15

After applying post-training quantization, the memory footprint of both MobileNetV1

and EfficientNet-B0 was significantly reduced, allowing the models to fit within the 256 KB
RAM of the Arduino Nano 33 BLE Sense while still maintaining high accuracy. The quan-
tized versions of both models continued to provide robust performance, making them
suitable for real-time gas leak detection on resource-constrained devices.
The inference time was measured at 1.4 s, well within the acceptable range for real-
time gas leak detection. The quantized models significantly reduced latency and power
consumption, which are crucial factors when deploying machine learning models on
low-power edge devices
Given the slightly higher accuracy and F1 scores of EfficientNet-B0 during testing, it
may be the preferred model for deployment in environments where higher detection accu-
racy is critical. However, MobileNetV1 remains a viable option, particularly in scenarios
where model simplicity and slightly lower computational overhead are prioritized.

5. Conclusions
This study highlights the critical importance of leveraging advanced techniques for
gas leak detection in residential and industrial settings. By integrating IoT and TinyML
technologies, we have developed a system that enhances the accuracy and efficiency of gas
leak detection while operating on resource-constrained devices.
Deploying our gas leak detection system on the Arduino Nano 33 BLE Sense, comple-
mented by the MLX90640 thermal camera, demonstrates the practical applicability of edge-
based intelligence. The combination of the ARM Cortex-M4F processor and MLX90640
thermal imaging facilitates detailed and real-time monitoring, crucial for effective edge-
based data analytics and immediate decision making.
Our experimental results show that the EfficientNet-B0 model achieved a testing accu-
racy of 91.73% with high recall rates, ensuring minimal false alarms and reliable detection
capabilities. The resource-efficient design of both the MobileNetV1 and EfficientNet-B0
models, optimized through post-training quantization, ensures seamless integration into
the Arduino Nano 33 BLE Sense, affirming our commitment to deploying cutting-edge
technology in practical applications.
The ability to process data locally using TinyML enhances system responsiveness and
reduces the likelihood of false alarms. This localized decision-making capability is crucial
for maintaining system efficiency and reliability, especially in environments where rapid
detection and response are essential.
The practical use of this gas leak detection system in actual settings will become the
main emphasis going forward. This will entail thorough testing, optimization, and fine-
tuning to guarantee dependability and efficacy in a range of scenarios. In future work, we
will focus on building the testbed and deploying the trained models on the Arduino Nano
33 BLE Sense to validate the system’s real-world performance. This deployment will allow
us to assess the models’ accuracy, energy efficiency, and responsiveness in real-time gas
leak detection.
In summary, the combination of IoT and TinyML technologies offers a reliable, ef-
fective, and expandable method for detecting gas leaks. The system’s potential for broad
use is demonstrated by the successful deployment of the MLX90640 thermal camera in
conjunction with the Arduino Nano 33 BLE Sense. This study highlights the revolutionary
effect of edge-based intelligence on environmental monitoring systems, opening the door
for future advancements in safety monitoring.
This study addressed the main research questions and demonstrated the effectiveness
of TinyML for gas leak detection in various settings:
• RQ1: TinyML models were found to enhance real-time detection, with EfficientNet B0
achieving an accuracy of 91.73% on low-power devices.
• RQ1: The balance between speed, processing load, and accuracy was managed ef-
fectively through techniques like post-quantization, which reduced model size while
maintaining reliable performance.
Electronics 2024, 13, 4768 14 of 15

• RQ1: Testing confirmed the strong performance of optimized models across different
conditions and hardware setups. MobileNetV1 provided an efficient, lightweight
option, while EfficientNet B0 delivered higher accuracy in settings with more avail-
able resources.
This work supports the use of TinyML for practical, scalable safety monitoring. Fu-
ture studies can expand upon this research by testing under more varied conditions and
exploring detection of different types of gases.

Author Contributions: Conceptualization, N.B. and H.T.; Methodology, M.E.B., N.B. and H.T.; Valida-
tion, M.E.B.; Formal analysis, M.E.B.; Investigation, M.E.B.; Data curation, M.E.B.; Writing—original
draft, M.E.B.; Writing—review and editing, N.B., H.T. and M.B; Supervision, N.B., H.T. and M.B.
All authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Data Availability Statement: The dataset described in this paper is available at https://data.
mendeley.com/datasets/zkwgkjkjn9/2 (accessed on 15 May 2024).
Conflicts of Interest: The authors declare no conflicts of interest.

References
1. Aman, F.; Puvanen Thiran, T.; Yusof, K.H.; Md Sapari, N. IoT Gas Leakage Detection Alert and Gas Concentration Reduction
System. In Proceedings of the 2022 IEEE 12th Symposium on Computer Applications & Industrial Electronics (ISCAIE), Penang,
Malaysia, 21–22 May 2022.
2. Lorthong, S.; Janjarassuk, U.; Jayranaiwachira, N. LPG Leakage Risk Predictions from an IoT-Based Detection System Using
Machine Learning. In Proceedings of the 2023 9th International Conference on Engineering, Applied Sciences, and Technology
(ICEAST), Vientiane, Laos, 1–4 June 2023.
3. Khan, M.M. Sensor-Based Gas Leakage Detector System. Eng. Proc. 2020, 2, 28. [CrossRef]
4. Abadade, Y.; Temouden, A.; Bamoumen, H.; Benamar, N.; Chtouki, Y.; Hafid, A.S. A Comprehensive Survey on TinyML. IEEE
Access 2023, 11, 3294111. [CrossRef]
5. Lin, J.; Zhu, L.; Chen, W.-M.; Wang, W.-C.; Han, S. Tiny Machine Learning: Progress and Futures. IEEE Circuits Syst. Mag. 2024,
23, 8–34. [CrossRef]
6. Pimpalkar, A.S.; Niture, D.V. Towards Contactless Elevators with TinyML using CNN-based Person Detection and Keyword
Spotting. IEEE Access 2024, 1–9. [CrossRef]
7. Atanane, O.; Mourhir, A.; Benamar, N.; Zennaro, M. Smart Buildings: Water Leakage Detection Using TinyML. Sensors 2023,
23, 9210. . [CrossRef] [PubMed]
8. Abadade, Y.; Benamar, N.; Bagaa, M.; Chaoui, H. Empowering Healthcare: TinyML for Precise Lung Disease Classification. Future
Internet 2024, 16, 391. [CrossRef]
9. Gkogkidis, A.; Tsoukas, V.; Papafotikas, S.; Boumpa, E.; Kakarountas, A. A TinyML-based system for gas leakage detection.
In Proceedings of the 11th International Conference on Modern Circuits and Systems Technologies (MOCAST), Bremen, Germany,
8–10 June 2022.
10. Kopbayeva, A.; Khan, F.; Yang, M.; Halim, S.Z. Gas leakage detection using spatial and temporal neural network model. Process
Saf. Environ. Prot. 2022, 159, 103–112. [CrossRef]
11. Miao, X.; Zhao, H.; Xiang, Z. Leakage detection in natural gas pipeline based on unsupervised learning and stress perception.
Process Saf. Environ. Prot. 2022, 164, 60–70. [CrossRef]
12. Doshmanziari, R.; Khaloozadeh, H.; Nikoofard, A. Gas pipeline leakage detection based on sensor fusion under model-based
fault detection framework. J. Pet. Sci. Eng. 2020, 184, 106581. [CrossRef]
13. Du, Y.; Jiang, J.; Yu, Z.; Liu, Z.; Pan, Y.; Xiong, K. A knowledge guided deep learning framework for underground natural gas
micro-leaks detection from hyperspectral imagery. Energy 2024, 294, 130847. [CrossRef]
14. Narkhede, C. Multimodal AI-based sensor fusion for gas detection. Sensors 2021, 21, 342–354.
15. Wang, J.; Liu, F.; Zhang, H.; Yang, X.; Yan, J. Invisible Gas Detection: An RGB-Thermal Cross Attention Network and A New
Benchmark. arXiv 2024, arXiv:2403.17712. [CrossRef]
16. Attallah, O.; Elhelw, A.M. Gas Leakage Recognition Using Manifold Convolutional Neural Networks and Infrared Thermal
Images. In Proceedings of the 2023 Congress in Computer Science, Computer Engineering, & Applied Computing (CSCE), Las
Vegas, NV, USA, 24–27 July 2023.
17. Sharma, A.; Khullar, V.; Kansal, I.; Chhabra, G.; Arora, P.; Popli, R.; Kumar, R. Gas Detection and Classification Using Multimodal
Data Based on Federated Learning. Sensors 2024, 24, 5904. [CrossRef] [PubMed]
18. Tsoukas, V.; Gkogkidis, A.; Boumpa, E.; Papafotikas, S.; Kakarountas, A. A Gas Leakage Detection Device Based on the Technology
of TinyML. Technologies 2023, 11, 45. [CrossRef]
Electronics 2024, 13, 4768 15 of 15

19. Shafin, S.S.; Karmakar, G.; Mareels, I. Obfuscated Memory Malware Detection in Resource-Constrained IoT Devices for Smart
City Applications. Sensors 2023, 23, 5348. [CrossRef] [PubMed]
20. Narkhede, P.; Walambe, R.; Chandel, P.; Mandaokar, S.; Kotecha, K. MultimodalGasData: Multimodal Dataset for Gas Detection
and Classification. Data 2022, 7, 112. [CrossRef]
21. Howard, A.G.; Zhu, M.; Chen, B.; Kalenichenko, D.; Wang, W.; Weyand, T.; Andreetto, M.; Adam, H. MobileNets: Efficient
Convolutional Neural Networks for Mobile Vision Applications. arXiv 2017, arXiv:1704.04861.
22. Tan, M.; Le, Q.V. EfficientNetV2: Smaller Models and Faster Training. Proc. Mach. Learn. Res. 2021, 139, 10096–10106.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.