Development of a smart wearable device for monitoring dog's health and behavior in real-time

MINISTRY OF EDUCATION AND TRAININGHO CHI MINH CITY UNIVERSITY OF TECHNOLOGY AND EDUCATION GRADUATION THESIS MAJOR: MECHATRONICS ENGINEERING TECHNOLOGY INSTRUCTOR: TRAN THUY UYEN PHUONG N

INTRODUCTION AND RESEARCH FOUNDATION

Introduction

In today's world, the connection between humans and pets, especially dogs, has become increasingly significant, making them beloved family members As pet ownership grows, ensuring their health and well-being is a top priority for responsible owners However, monitoring a pet's behavior and health can be challenging for those with busy schedules, leading to worries about their well-being and the risk of missing important changes that may need immediate attention.

The rising concern for pet well-being has led to the emergence of wearable pet monitoring technologies, offering a promising solution for pet owners These advanced devices utilize state-of-the-art sensors and data analysis to deliver insights into a pet's activities, behaviors, and health status, even in the owner's absence By monitoring physical activities and vital signs, pet owners can identify daily routines, detect early health issues, and make informed care decisions This proactive approach enables timely interventions, potentially preventing serious health problems and lowering veterinary expenses For veterinarians, access to continuous real-time data enhances diagnostic accuracy and allows for personalized treatment plans, ultimately improving health outcomes for dogs.

1.1.1 International Research Situation (Global Research Landscape)

The global rise of wearable devices for pet monitoring has attracted significant interest from both commercial sectors and academic institutions Companies are tapping into this market potential with innovative products like Fitbark, an activity and sleep monitor for dogs, and Whistle, a GPS and activity tracker that helps owners keep tabs on their pets' location and activity levels These devices utilize sensors such as accelerometers and GPS to monitor pets' movements, energy expenditure, and sleep patterns, offering essential insights into their overall health and well-being.

The academic research community has played a crucial role in advancing pet monitoring technologies, focusing on sensor data analysis, machine learning algorithms, and pattern recognition One study demonstrated the effectiveness of using accelerometer data to classify canine activities such as sitting, standing, and running, achieving high accuracy in behavior detection (Ollerenshaw, Barrett, & Robinson, 2020) Additionally, researchers developed a wearable device that combines accelerometers and physiological sensors to monitor dogs' heart rate variability and activity, offering a comprehensive view of pet health and enabling early detection of potential health issues (Barker, Amundin, & Lohilahti, 2019).

Recent advancements in pet monitoring technologies highlight the substantial progress in this field, providing pet owners and veterinarians with essential tools to enhance the health and well-being of their furry companions.

1.1.2 Domestic Research Situation (Local Research Contributions)

In Vietnam, the wearable devices market for pet monitoring is in its infancy, yet local researchers and tech enthusiasts are making significant strides A study by Dr Nguyen Thi Hoang at Vietnam National University Ho Chi Minh City focuses on delivering real-time data on pets' locations and activity levels, enhancing the understanding of this innovative field.

While promising initiatives exist, the adoption and commercialization of advanced pet care technologies in Vietnam are still in early stages However, with the increasing demand for innovative pet care solutions, it is expected that there will be a greater focus on developing and enhancing wearable devices that cater to the unique needs of Vietnamese pet owners.

Research Objectives

This study focuses on creating a smart wearable device designed to monitor a dog's health and behavior in real-time By offering pet owners valuable insights into their pets' well-being, the device empowers them to take proactive steps and seek veterinary care when needed.

The specific objectives of this research are as follows:

1 Design and develop a wearable hardware device integrated with various sensors, such as accelerometers, GPS, and temperature sensors, to collect data related to a dog's movements, location, and body temperature

2 Implement algorithms for processing and filtering sensor data to extract meaningful features and patterns

3 Develop a machine learning model capable of analyzing the processed sensor data to identify and classify the dog's behaviors and potential health issues

4 Integrate the wearable device with cloud computing services (AWS) for data storage, processing, and real-time monitoring

5 Create a user-friendly web application that displays the dog's behavior, health status, and other relevant information to the pet owner

6 Test and evaluate the overall system's performance, accuracy, and usability.

Research Methodology

This study aims to meet its research objectives through a multifaceted approach that includes hardware design, embedded systems development, algorithm implementation, machine learning model creation, cloud computing integration, and web application development The research will progress through several key steps to ensure comprehensive results.

1 Conduct a comprehensive literature review to understand the state-of-the-art technologies and methodologies related to wearable devices, pet monitoring, sensor technologies, machine learning, and cloud computing

2 Design and fabricate the hardware components, including the motherboard and the integration of various sensors

3 Develop firmware for the embedded system to collect and process sensor data, manage wireless communication, and monitor battery levels

4 Implement algorithms for data filtering, noise reduction, and feature extraction from the sensor data

5 Train and evaluate a machine learning model using the processed sensor data to classify the dog's behaviors and detect potential health issues

6 Set up the required cloud infrastructure on AWS, including IoT Core, Lambda functions, DynamoDB tables, and API Gateway

7 Deploy the trained machine learning model on an AWS EC2 instance for real-time behavior analysis

8 Develop a web application that integrates with the AWS services to display the dog's behavior, health status, and other relevant information to the pet owner

9 Conduct comprehensive testing and evaluation of the overall system, including hardware, firmware, cloud infrastructure, machine learning model, and web application

10 Document the research findings, conclusions, and potential future enhancements.

Scope of the Study

This project aims to create a real-time pet monitoring system for dogs, designed to enhance pet care by replacing the need for constant human supervision The system evaluates a dog's health and well-being, alerting owners to potential veterinary issues without diagnosing specific diseases By offering valuable health insights, it encourages proactive care and provides peace of mind for pet owners Currently in the academic research phase, further development is necessary before the system can be commercially available.

Data Collection: The system will gather comprehensive data on canine movement, temperature, and location Algorithms will be developed to facilitate basic behavioral and health monitoring functions

Connectivity: Leveraging both Wi-Fi and NB-IoT/LTE-M communication protocols, the device will ensure seamless compatibility with the existing Vietnamese network infrastructure

Applications: The system will be tailored for specific use cases:

- Veterinary Hospitals: Customized to enable continuous patient monitoring and data collection for research within veterinary settings

- Pet Stores: Adapted for demonstrative purposes, showcasing the potential of pet monitoring technology in retail environments

A comprehensive set of data collection and analysis tools will be created to enhance veterinary and animal behavior research in academic institutions, promoting a deeper understanding of canine health and behavioral patterns.

Thesis Organization

This thesis is organized into the following chapters:

Chapter 1: Introduction and Research Foundation

This chapter presents the research topic on pet monitoring technologies, offering essential background information and outlining the objectives and methodology of the study It further explores the global and local research landscape concerning wearable devices designed for pet monitoring.

This chapter outlines the theoretical framework for the project, focusing on wearable device technologies and various sensor technologies such as accelerometers, gyroscopes, GPS, and temperature sensors It also discusses microcontroller units and wireless communication technologies, alongside the roles of cloud computing and the Internet of Things (IoT) Additionally, it explores the application of machine learning for behavior analysis and the implementation of fuzzy logic systems.

Chapter 3: Hardware Design and Fabrication

This chapter details the hardware aspects of the project, including system requirements, component selection (sensors, NB-IoT module, GPS module, ESP32 microcontroller), motherboard design, power management, and component assembly

Chapter 4: Algorithms and Cloud Services

This chapter emphasizes the software and cloud components of the project, detailing the algorithms for processing sensor data, the development of machine learning models for behavior prediction, and the application of fuzzy logic for health forecasting Additionally, it discusses the deployment of cloud services utilizing AWS technologies, including IoT Core, Lambda, DynamoDB, and API Gateway.

Chapter 5: System Implementation and Testing

This chapter outlines the complete implementation of the system, detailing the integration of hardware and software components, the creation of the web application interface, and the thorough testing and evaluation of the system.

This chapter outlines the project results, highlighting the wearable device's performance, the accuracy of its behavior and health predictions, and the overall functionality of the system Additionally, it explores the implications of these findings and provides a comparison with current solutions in the market.

Chapter 7: Conclusion and Future Work

The concluding chapter highlights the project's significant accomplishments, addresses its limitations, and suggests avenues for future research and development in smart wearable technology for monitoring pets.

THEORETICAL BASIS

Wearable Device Technologies for Pet Monitoring

2.1.1 Overview of Current Pet Wearable Devices

Wearable devices have significantly transformed health monitoring and fitness tracking in recent years, and this innovative technology is now enhancing pet care Specialized wearable devices are enabling pet owners to monitor their animal companions' behavior and health, ushering in a new era of pet wellness.

The global pet wearable market has seen significant growth, with a projected value of

$3.48 billion by 2027 [6] This growth is driven by increasing pet ownership, rising concerns about pet health and safety, and advancements in sensor and communication technologies

Figure 2.1 Examples of Pet Wearable Devices and Applications for Health and Activity

Several products have gained prominence in the pet wearable market, each offering unique features:

1 Whistle GO Explore: A GPS pet tracker that provides real-time location tracking, activity monitoring, and health trend analysis This device uses advanced GPS and cellular technology to offer precise location tracking within 10 feet It also monitors scratching, licking, and sleeping patterns, providing insights into potential health issues The device is waterproof and has a battery life of up to 20 days [1]

2 FitBark 2: Focuses on activity tracking, sleep quality monitoring, and calorie expenditure estimation This lightweight device (10 grams) can be attached to any collar and provides 24/7 monitoring It tracks activity levels, distance traveled, calories burned, sleep quality, and overall health and behavior The FitBark 2 is waterproof

3 PetPace: Offers continuous monitoring of vital signs, including temperature, pulse, respiration, activity patterns, calories, pain, and various behavioral parameters This collar uses non-invasive sensors to track vital signs and transmit them to a cloud-based engine for analysis [8] It can alert owners and veterinarians to potential health issues before they become serious The device is particularly useful for monitoring pets with chronic conditions or those recovering from surgery

These devices are typically designed to be small, lightweight, and easily attachable to a pet's collar or harness The key hardware components in most pet wearable devices include:

1 Microcontroller: Acts as the central processing unit, managing data collection, processing, and transmission

• Accelerometers: Measure acceleration forces, detecting the pet's movements and activity levels

• Gyroscopes: Measure angular velocity, providing more accurate data on the pet's orientation and movement patterns

• Temperature sensors: Monitor the pet's body temperature

• Heart rate monitors: Some advanced devices use photoplethysmography (PPG) sensors for non-invasive heart rate monitoring

• Microphones: Detect barking patterns or other vocalizations, indicating the pet's emotional state or potential issues

• GPS module: For location tracking and geofencing capabilities

3 Wireless communication module: Often utilizing Bluetooth, Wi-Fi, or cellular networks (3G/4G/5G)

4 Battery: Providing power to the device, with life varying from days to weeks depending on usage and features

The most measured parameters in pet wearable devices include:

• Activity levels: Measured through step counts and motion intensity, helping owners ensure their pets get adequate exercise

• Sleep patterns: Monitoring rest periods can indicate overall health and stress levels

• Location: GPS tracking for safety and to prevent loss

• Temperature: Body temperature can be an indicator of health issues

• Heart rate: In some advanced devices, are used to monitor cardiovascular health

• Respiratory rate: Can indicate stress levels or potential respiratory issues

• Calories burned: Helps in maintaining a healthy weight for pets

These parameters are measured for several important reasons:

1 Early detection of health issues: Changes in activity levels, sleep patterns, or vital signs can indicate potential health problems before they become severe

2 Behavior understanding: Tracking daily routines helps owners better understand their pet's needs and habits

3 Weight management: Activity tracking and calorie estimation aid in maintaining a healthy weight, and preventing obesity-related health issues

4 Safety: Location tracking helps prevent pets from getting lost and enables quick recovery if they do wander off

5 Veterinary care support: Continuous monitoring provides veterinarians with valuable data for more accurate diagnoses and treatment plans

Wearable devices for pet monitoring offer significant advantages, primarily through their ability to wirelessly collect and transmit data Utilizing technologies like Bluetooth Low Energy (BLE) for short-range communication and Wi-Fi or cellular networks for longer distances, these devices allow pet owners to access important information remotely This connectivity empowers owners to monitor their pets' activities and well-being from anywhere, even when they are not physically present.

Advanced pet wearable devices utilize cloud computing and Internet of Things (IoT) technologies to improve functionality and data processing By sending collected data to cloud servers, these devices harness powerful computing resources for advanced data analysis, pattern recognition, and machine learning algorithms, enabling the identification of potential health issues and behavioral patterns in pets.

The fusion of wearable devices with cloud computing and IoT technologies has revolutionized pet monitoring, allowing pet owners to receive instant alerts on their smartphones for any unusual behavior or health issues detected by the system Furthermore, veterinarians and pet care professionals can remotely access this valuable data, facilitating personalized and informed care for pets.

Despite the numerous benefits and potential applications of wearable devices for pet monitoring, there are still challenges and limitations that need to be addressed These include:

• Accuracy and reliability of sensor data, which can be affected by factors such as pet movement patterns and environmental conditions

• Battery life limitations, especially for devices with continuous GPS tracking

• Data privacy and security concerns, as these devices collect sensitive information about pets and their owners

• User adoption and ease of use, particularly for less tech-savvy pet owners

• Durability and water resistance, as pets can be rough on devices and often encounter water

Ongoing research and development in wearable devices for pet monitoring is expected to overcome current challenges, resulting in more accurate, reliable, and user-friendly solutions This rapidly evolving field is fueled by advancements in sensor technologies and data analytics, alongside a growing demand for comprehensive pet care Future innovations may encompass sophisticated health monitoring features, extended battery life, enhanced data security, and improved integration with veterinary care systems.

Wearable device technologies for pet monitoring mark a major advancement in pet care, providing pet owners with valuable insights into their pets' health and behavior As these technologies evolve and address current challenges, they hold the potential to transform our approach to pet care, enhancing health outcomes and strengthening the bond between humans and animals.

Sensor Technologies

Advanced sensor technologies, including accelerometers and gyroscopes, are essential for effectively monitoring and diagnosing a dog's health and activity levels These sensors accurately track a dog's movements and orientation, offering valuable insights into their behavior and well-being Additionally, the integration of GPS modules and temperature sensors enhances this monitoring capability, providing a comprehensive overview of a dog's overall health status.

Figure 2.2 Sensing Principles of Accelerometers and Gyroscopes in Motion Tracking[27]

Dog tracking systems rely on accelerometers and gyroscopes to deliver detailed insights into a dog's movement and health Accelerometers track changes in velocity, while gyroscopes measure angular velocity, together offering crucial data on a dog's activity levels and orientation This information, when processed, can significantly enhance understanding of a dog's overall well-being.

Raw sensor data from accelerometers and gyroscopes is analyzed to extract key metrics such as activity intensity, posture recognition, step counting, abnormal vibration detection, and estimated energy consumption Activity intensity is derived from the three-dimensional acceleration vector, averaged over minute intervals By integrating accelerometer tilt angles with gyroscope angular velocity, the orientation can be accurately identified Step counting relies on the accelerometer to detect peaks in longitudinal acceleration, with gyroscope data enhancing directional accuracy Abnormal vibrations are identified through high-frequency data analysis from both sensors Estimated energy consumption is calculated mainly from accelerometer data, with adjustments from gyroscopic data for rotational motion These metrics are obtained through on-device processing and advanced server analytics, providing a thorough evaluation of a dog's activity patterns and overall health.

Feature extraction and filtering utilize a blend of on-device signal processing and sophisticated server-based analytics By combining data from accelerometers and gyroscopes, dog tracking systems deliver a more precise and thorough evaluation of canine activity and health compared to relying on a single sensor type alone.

GPS modules in dog monitoring devices are essential for tracking canine activities outdoors, providing vital location data that enhances our understanding of dogs' movements and behaviors Although their effectiveness is limited to areas with clear sky visibility, when combined with inertial sensors, GPS technology allows for a detailed analysis of dogs' outdoor routines, exercise habits, and interactions with their surroundings.

The GPS module offers essential insights into a dog's outdoor activities by providing precise location tracking through absolute position data, including latitude and longitude It calculates movement metrics such as speed and total distance traveled based on position changes over time By analyzing clustered GPS data points, owners can identify activity zones, revealing frequently visited locations and the duration spent in various areas Furthermore, escape detection features enhance safety by alerting owners to rapid position changes or when the dog exits permitted zones.

Extracting and analyzing GPS-based features offers essential insights into a dog's outdoor activities and environmental exposures When this location-specific data is integrated with information from other sensors, it creates a comprehensive understanding of the dog's daily routines and potential health indicators This thorough approach to monitoring dogs facilitates informed decisions regarding pet care and overall health management.

Temperature sensors in dog monitoring devices are essential for evaluating a dog's health and environmental conditions Typically affixed to the collar or harness, these sensors make contact with the dog's skin in areas with minimal fur, like the neck or chest They function at low frequencies, taking measurements every few minutes to ensure a balance between accuracy and power efficiency.

Temperature sensors play a crucial role in monitoring a dog's health by providing essential data on their physiological state and environmental conditions The key feature is the absolute temperature reading, which helps identify fever or hypothermia Additionally, temporal temperature gradients reveal changes in temperature over time, indicating potential illness or variations in activity levels Temperature variability, assessed through the range and standard deviation of readings, can uncover patterns related to circadian rhythms or unusual fluctuations To ensure accurate data, various filtering techniques are utilized, including moving average filters to reduce short-term noise, outlier detection algorithms to eliminate erroneous readings, and adaptive thresholding to accommodate breed-specific normal temperature ranges.

Analyzing temperature-based features provides essential insights into a dog's physiological condition and environmental interactions When combined with data from additional sensors, this information creates a holistic view of the dog's health and well-being Consequently, the temperature sensor plays a vital role in the ecosystem of canine health monitoring technologies, facilitating early detection of health problems and informed care decisions for the dog.

Microcontroller Unit

Microcontroller units (MCUs) play a vital role in wearable pet monitoring devices, serving as the central processing hub that integrates a processor, memory, and I/O peripherals on a single chip This compact design is ideal for embedded applications where space and power efficiency are essential In pet wearables, MCUs function as the device's "brain," coordinating real-time data collection from various sensors like accelerometers, gyroscopes, and GPS modules, while also managing wireless communication for data transmission.

The perfect MCU for pet wearables should be compact and lightweight to ensure animal comfort, while providing adequate processing power for real-time data acquisition and analysis from various sensors It's essential to prioritize energy efficiency to extend battery life and minimize the need for frequent recharging Additionally, robust connectivity options, especially support for wireless protocols like Bluetooth and Wi-Fi, are crucial for seamless communication.

Fi, are critical for enabling remote monitoring and data transmission

When choosing the best MCU for pet monitoring applications, a comparative analysis of popular options such as the ESP32, ATmega series, Raspberry Pi Zero, and STM32 series highlights the unique advantages and limitations of each platform.

The ESP32 is perfect for applications requiring robust wireless connectivity, thanks to its built-in Wi-Fi and Bluetooth capabilities With a dual-core processor, it efficiently processes real-time data from various sensors, striking a balance between processing power and memory at an affordable price.

The ATmega series offers simplicity, efficiency, and user-friendly programming, backed by robust community support With low power consumption, it is ideal for basic monitoring applications; however, it does not include built-in wireless connectivity, necessitating additional modules that can add complexity and increase costs.

The Raspberry Pi Zero is compatible with a range of peripherals and sensors, including Wi-Fi and Bluetooth While it is capable of handling complex data processing tasks, its higher power consumption and larger size make it less suitable for compact, battery-powered wearable devices.

The STM32 series, particularly the STM32F4 models, provides an excellent blend of energy efficiency and computational power through ARM Cortex-M processors These features make it ideal for battery-operated applications, although it may involve increased programming complexity and higher costs.

Advantages of ESP32 Compared to Alternatives

• Versus STM32: The ESP32 offers integrated wireless capabilities and generally superior RAM and flash memory at comparable price points [10]

• Versus Raspberry Pi: It provides significantly lower power consumption and a smaller form factor, crucial for our wearable application [11]

• Versus Arduino: The ESP32 boasts superior processing power, memory capacity, and built-in wireless connectivity.

Wireless Communication Technologies

2.4.1 Introduction to Wireless Communication Technologies

Wireless communication technologies are essential for the seamless interaction and data exchange between devices, serving as the foundation of the Internet of Things (IoT) In our innovative smart wearable device for monitoring dogs' health and behavior in real-time, these technologies are crucial for enabling continuous data transmission and analysis.

Our project utilizes Wi-Fi and Narrowband Internet of Things (NB-IoT) to create a versatile communication infrastructure This combination enhances connectivity across various environments, from urban areas with extensive Wi-Fi coverage to remote locations requiring low-power wide-area network (LPWAN) solutions like NB-IoT for dependable long-range communication.

Wi-Fi, or Wireless Fidelity, is a collection of wireless network protocols based on IEEE 802.11 standards, utilizing radio waves for high-speed internet connections primarily in the 2.4 GHz and 5 GHz frequency bands Key components of Wi-Fi networks include Access Points (APs) that broadcast signals, routers that manage local and internet traffic, and client devices, such as smart wearable devices for dogs This technology enables efficient data transmission within local networks, providing essential communication for real-time monitoring of pet health and behavior.

Figure 2.3 Ecosystem of Wi-Fi Connected Devices and Applications[9]

The Wi-Fi communication process involves the following steps:

The access point broadcasts its SSID (Service Set Identifier)

1 The client device detects this signal and initiates a connection request

2 If security is enabled (e.g., WPA2), the devices perform a handshake to establish a secure connection [6]

3 Once connected, the client device can send and receive data through the access point

Wi-Fi uses a technique called Orthogonal Frequency-Division Multiplexing (OFDM) to transmit data efficiently, allowing for high data rates even in environments with potential interference

1 High Data Transmission Rates: Modern Wi-Fi standards like 802.11ac can achieve theoretical speeds up to 3.46 Gbps, though real-world speeds are typically lower This high bandwidth is crucial for transmitting large volumes of sensor data quickly

2 Low Operational Costs: Once a Wi-Fi network is set up, data transmission is essentially free, making it a cost-effective option for continuous monitoring

3 Widespread Availability: Wi-Fi networks are ubiquitous in urban and residential areas, providing ample connectivity options for our device

4 Lower Power Consumption: Compared to cellular technologies, Wi-Fi generally consumes less power, which is critical for extending the battery life of our wearable device

5 Easy Integration: Many microcontrollers, including the ESP32 used in our project, come with built-in Wi-Fi capabilities, simplifying the hardware design

1 Limited Range: Wi-Fi signals typically have an effective range of 30-50 meters indoors, which can be a constraint for monitoring dogs in large outdoor areas

2 Network Dependency: The device relies on the availability and accessibility of known Wi-Fi networks, which may not always be present in all locations a dog might visit

Narrowband Internet of Things (NB-IoT) is a Low Power Wide Area Network (LPWAN) technology designed for a variety of cellular devices and services It offers exceptional indoor coverage, cost-effective implementation, extended battery life, and high connection density, making it perfect for pet monitoring systems These advantages are essential for tracking dogs in diverse environments, ensuring reliable connectivity, and allowing devices to operate for long periods without the need for frequent recharging By leveraging existing cellular infrastructure, including base stations and core networks, NB-IoT seamlessly integrates into devices like smart wearables for dogs.

The communication process utilizes device registration, authentication, and narrowband signal transmission to ensure reliable data transfer, even in challenging environments This method enhances range and barrier penetration, allowing for effective monitoring of pets indoors By employing NB-IoT technology, our system enables continuous tracking of pet health and behavior while maintaining low power consumption, making it a cost-effective solution for long-term use.

Figure 2.4 Applications of Narrowband IoT (NB-IoT) Across Different Domains[13]

1 Extended Coverage: NB-IoT provides wide coverage

2 Low Power Consumption: Designed for IoT devices, NB-IoT allows for years of battery life on a single charge [13]

3 Wide Area Coverage: Can transmit data over much longer distances compared to Wi-

Fi, ideal for tracking dogs in large or remote areas

4 High Connection Density: Supports a massive number of connected devices per cell

5 Licensed Spectrum: Operates in licensed bands, ensuring better reliability and less interference

1 Lower Data Rates: NB-IoT is designed for small, infrequent data transmissions, with speeds typically up to 250 kbps

2 Higher Latency: Compared to Wi-Fi, NB-IoT can have higher latency, which may affect real-time applications

3 Network Availability: Depends on cellular network coverage and NB-IoT deployment in the area

4 Subscription Costs: Requires a cellular data plan, albeit at lower costs compared to traditional cellular services.

Cloud Computing and Internet of Things

The rise of cloud computing and the Internet of Things (IoT) has transformed data collection and analysis from various sources, particularly wearable devices These technologies enable real-time monitoring and efficient analysis of sensor data, facilitating scalable solutions in diverse areas like pet monitoring Cloud computing provides on-demand access to computing services, such as storage and processing power, over the Internet, eliminating the need for local infrastructure This shift has revolutionized application development and maintenance, offering significant benefits, including cost-effectiveness, flexibility, and scalability.

Figure 2.5 Cloud-Based Architectures for Internet of Things (IoT) Devices [15]

Amazon Web Services (AWS) is a leading provider of cloud computing services, offering a wide range of solutions including computing, storage, databases, and analytics These services empower developers and businesses to quickly and securely build and deploy applications In the realm of pet monitoring through wearable devices, cloud computing is crucial for system architecture, as it ensures the secure transmission and storage of data collected from these devices AWS provides essential tools like AWS IoT Core and Amazon Kinesis, which support secure and scalable data ingestion from IoT devices.

Once data is stored in the cloud, it can be efficiently processed and analyzed using various AWS services Amazon Lambda, a serverless computing solution, allows for code execution without the need to manage servers, facilitating cost-effective data processing Lambda functions can be triggered by events like the arrival of new data, enabling real-time analysis of sensor information The analyzed data can then be stored in Amazon DynamoDB, a fully managed NoSQL database that ensures durable data persistence and easy access for web and mobile applications.

AWS provides powerful machine learning services like Amazon SageMaker, which facilitates the building, training, and deployment of machine learning models for advanced data analysis and behavior prediction These models utilize processed sensor data and are deployed on AWS infrastructure, allowing for real-time inference and predictions as new data streams in from wearable devices The Internet of Things (IoT) is essential for seamless communication and data exchange between these devices and cloud services AWS IoT Core offers a secure and scalable platform for managing IoT devices, ensuring secure communication, effective data processing, and smooth integration with other AWS services.

The integration of cloud computing and the Internet of Things in pet monitoring systems harnesses the scalability and reliability of AWS services This combination enables real-time monitoring, efficient data processing, and advanced analytics, including machine learning models for behavior and health analysis As a result, pet owners gain essential insights into their pets' well-being, enhancing their ability to ensure optimal care.

AWS provides various services to enhance the functionality and user experience of pet monitoring systems Amazon API Gateway facilitates the creation and management of APIs for seamless integration with web and mobile applications Amazon Cognito ensures secure user authentication and authorization for accessing system resources Additionally, Amazon CloudFront, a content delivery network (CDN), efficiently distributes static content like web pages and media files, thereby improving overall performance and user satisfaction.

Machine Learning for Behavior Analysis

2.6.1 Recent Machine Learning Techniques Applied to Animal Behavior Analysis

The rapid advancement of artificial intelligence has led to its widespread applications across various fields, significantly enhancing our quality of life Notably, machine learning has transformed animal behavior analysis by offering powerful tools for interpreting complex data patterns Recent years have seen the implementation of diverse machine-learning techniques to monitor and analyze animal behavior, particularly in dogs, resulting in remarkable improvements in both accuracy and efficiency.

A significant study by Chen et al employed supervised learning algorithms, including decision trees and support vector machines, to classify dairy cow behaviors using accelerometer data, achieving over 90% accuracy in identifying activities like lying, standing, and walking In a similar vein, Dogan et al utilized neural networks to classify dog behaviors from wearable sensor data, reaching an impressive accuracy rate of 92%.

Unsupervised learning techniques, such as clustering and anomaly detection, are essential for detecting unusual behaviors that may signal health concerns A notable study by Martiskainen et al utilized k-means clustering on cattle movement data, effectively pinpointing deviations from typical behavior patterns that could indicate potential health issues.

Deep learning techniques, especially convolutional neural networks (CNNs) and recurrent neural networks (RNNs), have excelled in processing complex, high-dimensional data like video and audio For instance, a study by Kim et al achieved an impressive 95% classification accuracy of various behavioral states in laboratory mice using CNNs.

[21] employed RNNs to analyze audio data for identifying stress-related vocalizations in pigs, achieving an accuracy of 89%

Recent research highlights the effectiveness of machine learning and artificial intelligence in analyzing animal behavior, demonstrating their high efficiency The integration of these technologies with wearable devices, such as the one proposed in our project, allows for continuous, real-time monitoring of a dog's behavior and health This synergy not only improves the accuracy of behavior prediction and health assessment but also provides scalable solutions for pet owners and veterinarians Furthermore, advanced machine learning models can identify early signs of distress or illness, enabling timely interventions that significantly enhance animal welfare and health outcomes.

In our project, we utilize two different machine learning algorithms, Support Vector Machine (SVM) and Long Short-Term Memory (LSTM) networks, to analyze and predict dog behaviors This is achieved by training the models with data collected from gyroscope and accelerometer sensors mounted on the dog's neck and back.

We utilize both algorithms for their unique strengths in addressing various data aspects Following extensive testing and assessment, we will choose the model that exhibits the greatest effectiveness and accuracy This article will explore the fundamental principles behind each algorithm, emphasizing their advantages By doing so, we aim to showcase how these strengths enhance the overall performance of our canine behavior prediction system.

Support Vector Machine (SVM) is a widely used supervised learning algorithm primarily employed for classification tasks in machine learning, though it can also handle regression problems The main objective of SVM is to establish an optimal decision boundary, known as a hyperplane, that effectively separates n-dimensional space into distinct classes This hyperplane is determined by maximizing the margin, which is the greatest distance between data points of the different classes Understanding these concepts is essential for leveraging SVM in various applications.

Figure 2.6 Description of SVM algorithm [22]

Hyperplane and Support Vectors in the SVM algorithm:

In n-dimensional space, various lines or boundaries can distinguish different classes of data The objective is to determine the optimal boundary that effectively classifies these data points, referred to as the hyperplane in a Support Vector Machine.

The hyperplane's shape is determined by the number of features in a dataset; it appears as a straight line with two features and as a two-dimensional plane with three features Our objective is to construct a hyperplane that maximizes the margin, defined as the distance between the hyperplane and the closest data points from each class.

Support Vectors are the data points closest to the hyperplane that influence its positioning These crucial vectors, which play a key role in defining the hyperplane, are termed Support Vectors due to their supportive function.

There are two types of support vector machines:

Figure 2.7 Example for Linear SVM[23]

Linear SVM is designed for datasets that are linearly separable, meaning that two classes can be distinctly classified by a single straight line This classifier effectively distinguishes between these two classes using the principles of linear separation.

Non-linear SVM is a powerful classification technique used for datasets that cannot be separated by a straight line When data is non-linearly separable, the Non-linear SVM classifier is employed to effectively categorize the information.

We selected Support Vector Machine (SVM) for predicting dog behaviors due to its effectiveness in high-dimensional spaces and its capability to establish clear margins between various behavior classes Its simplicity and low computational complexity make SVM particularly suitable for deployment on AWS EC2, which typically offers less computational power than local machines Additionally, research findings by Chen et al [17] have highlighted the significant potential of the SVM algorithm in this context.

2.6.2.2 Long Short-Term Memory (LSTM)

Long Short-Term Memory Networks (LSTMs) are a specialized type of Recurrent Neural Network (RNN) designed to learn long-term dependencies Introduced by Hochreiter and Schmidhuber in 1997, LSTMs have been refined and popularized by subsequent research They excel in handling a variety of problems, particularly in time-series data, due to their ability to learn and remember over extended sequences LSTMs effectively address many limitations of traditional RNNs, although their architecture is more complex.

A standard RNN network will have a very simple architecture such as the architecture consisting of one hidden layer which is the tanh function as shown below

Figure 2.8 The repeating module in a standard RNN contains a single layer [23]

Fuzzy Logic Systems

Fuzzy logic is a mathematical framework that simulates human decision-making by accommodating degrees of truth, unlike traditional boolean logic, which only considers binary values of true or false Introduced by Lotfi Zadeh in 1965, fuzzy logic effectively addresses imprecise and uncertain information, making it valuable in diverse applications such as control systems, artificial intelligence, and data analysis.

Fuzzy Sets and Membership Functions

Fuzzy logic is fundamentally based on fuzzy sets, which allow elements to partially belong to a set, contrasting with classical set theory The degree of membership, varying from 0 to 1, is determined by membership functions that assign a corresponding degree to each point in the input space Common types of membership functions include triangular, trapezoidal, Gaussian, and sigmoidal, with the selection of a specific function depending on the application and the data being modeled.

Figure 2.16 Block diagram of a fuzzy logic system

A fuzzy inference system (FIS) utilizes fuzzy logic to convert inputs into outputs through a structured process It involves fuzzification, which transforms crisp inputs into fuzzy values, and a rule base that outlines the input-output relationships The inference engine then applies these rules to the fuzzified inputs, followed by defuzzification, which reverts the fuzzy output to a crisp value There are two primary types of FIS: Mamdani, which employs fuzzy sets as outputs, and Sugeno, which uses linear functions of the inputs.

Fuzzy logic is highly effective in health monitoring and diagnosis, as it adeptly manages imprecise information and simulates human reasoning By incorporating expert knowledge through fuzzy rules, it effectively addresses uncertainties in health data, resulting in interpretable outcomes for healthcare professionals In the realm of pet health monitoring, fuzzy logic proves invaluable by analyzing various inputs, such as temperature and activity levels, to detect behavioral or physiological anomalies and offer personalized health recommendations based on fuzzy rules.

Fuzzy Logic in Wearable Devices

Fuzzy logic enhances wearable health monitoring devices by providing efficient processing and adaptability, making it suitable for resource-limited environments It enables personalized health tracking through the integration of new data and expert insights The interpretability of fuzzy systems, characterized by rules and membership functions that reflect human understanding of health metrics, promotes transparent decision-making Furthermore, fuzzy logic excels in managing noisy or incomplete sensor data, which is crucial for real-world applications These attributes position fuzzy logic as a vital component in creating intelligent health monitoring systems for wearables, especially in pet health monitoring where exact numerical thresholds may be lacking.

HARDWARE DESIGN AND FABRICATION

System Requirements

The main goal of our pet monitoring system's hardware is to effectively gather detailed data and transmit it securely to a central server, all while prioritizing the pet's comfort and safety To meet this objective, we have outlined specific system requirements.

- Dual-sensor configuration: One sensor on the dog's back and another on the ventral neck area

- Three primary data categories: Movement information, core body temperature, and position

- Advanced motion sensors: Accelerometer (±16g range) and gyroscope (±2000 degrees/second range)

- Precision temperature monitoring: 30°C to 45°C range, ±0.1°C accuracy, sampled once per second

- High-precision geolocation: GPS accuracy within 5 meters sampled once per second Device Specifications: Compact design:

- Primary unit (back-mounted): Max 9cm x 8cm x 4cm

- Secondary unit (neck-mounted): Max 3cm x 2cm x 2cm

- Battery life: Minimum 12 hours of continuous operation

- Durability: Operating temperature 0°C to 50°C, able to withstand practical impact forces (1-meter drop test, pressure from a dog on the device)

Data Transmission: Dual-mode communication:

- Wi-Fi for proximal data transfer

- LTE-M or NB-IoT for extended-range transmission to the central server

To ensure reliable data collection across diverse dog activities and environments while prioritizing pet comfort, our hardware must meet specific requirements This thorough approach to data gathering and transmission poses substantial engineering challenges, yet it is crucial for developing a robust and effective pet monitoring system.

Hardware Design

Based on the outlined requirements, our proposed design comprises two components:

- Primary unit (dog's back): Houses the MPU9255, temperature sensor, GPS module, ESP32 microcontroller, battery, and NB-IoT module for data processing and transmission

- Secondary unit (dog's neck): A compact, lightweight housing for a second MPU6050 sensor, connected to the primary unit for data collection and transmission

3.2.1.1 Acceleration and gyroscope sensor: MPU9255 and MPU6050

When developing a sophisticated canine behavior monitoring device, selecting the right sensor technology is crucial for its efficacy and reliability The MPU6050 and MPU9255 sensors are preferred for their integration of 3-axis accelerometers and gyroscopes, enabling advanced sensor fusion algorithms for precise motion tracking With measuring ranges of ±16g for accelerometers and ±2000°/s for gyroscopes, they effectively capture a wide range of canine movements These IMUs are durable and reliable, performing well in various conditions, including vibrations and impacts Their compact sizes (MPU6050: 20x16x2mm, MPU9255: 31.11x16.84x2.1mm) ensure minimal discomfort for the dog, making them ideal for a non-intrusive monitoring solution Overall, these features provide an optimal balance of accurate motion sensing, power efficiency, and practical wearability for canine behavior monitoring.

In designing our dog behavior monitoring device, selecting the appropriate NB-IoT module was crucial to ensure reliable data transmission while maintaining power efficiency

We evaluated three popular options available in Vietnam: SIM7080G, SIM7600, and

SIM7020 Each module offers distinct features and capabilities, which we assessed based on our specific requirements for size, power consumption, data transmission speed, and compatibility with our cloud infrastructure

The SIM7080G is a versatile module that supports both NB-IoT and CAT-M, offering uplink speeds of 150 Kbps and 1119 Kbps, respectively Its compact dimensions (65mm x 30.5mm) and low power consumption make it ideal for wearable devices Key features include TLS/SSL encryption, MQTT protocol compatibility, and easy integration with AWS IoT Core While the SIM7600 offers higher data speeds and a broader feature set, its larger size, higher cost, and greater power consumption make it less suitable for applications where device size and battery life are critical The SIM7020, similar in size and efficiency to the SIM7080G, supports NB-IoT but faces challenges with limited AT command length, hindering secure connections to AWS IoT Core, which is essential for cloud-based systems.

The SIM7080G was chosen as the ideal NB-IoT module for our dog behavior monitoring system due to its compact size, low power consumption, and adequate data rate, along with seamless integration with AWS IoT Core This selection enabled the creation of an efficient and reliable wearable device that continuously transmits data to our cloud platform.

Calculating the link budget is essential for assessing the performance and range of our NB-IoT-based dog behavior monitoring device For the SIM7080G module, we take into account several important factors: a transmit power (PTX) of 23 dBm, which is standard for NB-IoT devices; an antenna gain (GTX) of 0 dBi, reflecting the use of an omnidirectional antenna; a path loss (PL) of 140 dB, typical in urban environments; and a receiver sensitivity (SRX) of -129 dBm, which is common for NB-IoT base stations.

Link Budget = PTX + GTX - PL + SRX (3.1) Link Budget = 23 dBm + 0 dBi - 140 dB + 129 dBm = 12 dB

This positive link budget of 12 dB indicates a strong margin for reliable communication

Our device is designed to enhance connectivity in challenging environments by allowing for additional penetration through walls and obstacles (up to 20 dB), accommodating fading margins (around 10 dB), and compensating for body loss when worn by a dog (approximately 3-5 dB).

The SIM7080G module boasts an impressive maximum coupling loss (MCL) of 164 dB, enhancing its performance in challenging conditions This capability allows for an extended range of connectivity, reaching 10-15 km in rural areas and 1-2 km in dense urban settings Furthermore, it ensures deep indoor penetration, maintaining reliable connections even in basements or through multiple walls Overall, the SIM7080G provides robust connectivity for dog behavior monitoring devices, facilitating dependable data transmission in various environments, including those with poor traditional cellular coverage.

The integration of the SIM7080 module begins with provisioning a Viettel M2M SIM card and verifying it through the AT command "AT+CPIN?" After configuring the module for NB-IoT mode, it registers to the Viettel network using "AT+COPS=1,2, "45204"" Network connectivity is validated by checking the bearer activation status with "AT+CGACT?" and pinging a DNS server For data transfer to AWS IoT Core, TLS/SSL settings are established, including the MQTT server URL, port, client ID, and QoS level, while specifying the SSL/TLS version and converting certificates to a compatible format Finally, SSL/TLS is enabled for the socket connection, and data is published in JSON format to a designated MQTT topic using AT commands.

Secure data transmission was established through meticulous configuration of TLS/SSL protocols and certificate management, ensuring the continuous flow of behavior data to our cloud platform for analysis

3.2.1.3 GPS (Global Positioning System): SAM-M8Q

We chose the SAM-M8Q GPS module for our canine behavior monitoring device due to its exceptional performance and compatibility with our application needs Its multi-GNSS support, which allows for simultaneous reception of GPS, GLONASS, and Galileo signals, significantly improves positional accuracy and coverage, ensuring reliable tracking in various environments.

The module offers outstanding sensitivity with tracking capabilities down to -167 dBm and acquisition at -148 dBm, ensuring reliable performance in difficult signal environments like urban canyons and dense foliage This feature is crucial for our application, where reliable tracking is essential Additionally, the module provides data output in the standard NMEA format.

(National Marine Electronics Association) 0183 format, providing a wealth of information essential for our application[32]:

- GGA (Global Positioning System Fix Data): Delivers precise time, latitude, longitude, fix quality, and the number of satellites used for positioning

- RMC (Recommended Minimum Specific GPS/Transit Data): Includes time, status, latitude, longitude, speed, and course, offering a concise snapshot of the dog's movement

- GSV (Satellites in View): Provides detailed information on satellites in view, aiding in understanding signal quality and availability

This standardized data stream allows for accurate capture of location, speed, and time information, essential for detailed behavioral analysis The module's low power consumption of 20 mA during tracking and its compact size of 15.5 x 15.5 x 6.3 mm enhance the wearable design of our device, optimizing battery life and reducing bulk Additionally, the quick Time-To-First-Fix feature guarantees swift location acquisition upon activation.

The ESP32-WROOM-32U was chosen for our canine behavior monitoring device due to its outstanding performance, functionality, and cost-effectiveness Its dual-core architecture with two Xtensa 32-bit LX6 microprocessors running at up to 240 MHz ensures robust processing power for complex data analysis The integrated Wi-Fi (802.11 b/g/n) and Bluetooth 4.2 capabilities streamline the design by eliminating the need for additional wireless modules With various power modes to optimize battery life for wearable applications and a rich array of I/O options for sensor integration, the module is highly versatile Additionally, the U.FL connector for an external antenna enhances signal range and system reliability Overall, the ESP32-WROOM-32U provides an excellent performance-to-price ratio, allowing for more compact and economical motherboard designs.

The ESP32 microcontroller, paired with the ESPAsyncWebServer and AsyncElegantOTA libraries, offers significant benefits for connectivity and updates in pet monitoring systems With the ESPAsyncWebServer library, users can utilize a Wi-Fi Manager to effortlessly adjust device settings and toggle between Wi-Fi and NB-IoT modes via an intuitive web interface, removing the necessity for physical access This streamlined approach significantly improves user experience, enabling quick and hassle-free device setup.

The AsyncElegantOTA library enhances pet monitoring systems by providing a reliable over-the-air (OTA) update mechanism through a dedicated web server, allowing for remote firmware, sketch, and filesystem updates without physical connections This functionality is crucial for research and commercial applications, facilitating remote maintenance of deployed devices The compatibility of AsyncElegantOTA with other libraries ensures smooth integration of OTA features into web server projects, making the system more flexible, maintainable, and user-friendly.

Our pet monitoring system features a compact motherboard that acts as a versatile docking station for integrated sensors like the MPU9255 and MPU6050 Designed for space optimization and efficient connectivity, the board layout ensures peak performance within a small footprint The development process involves three key steps: schematic design, component placement, and trace routing.

Figure 3.6 Schematic Overview The schematic design incorporates the following key components:

1 Battery: Power source for the entire system

Figure 3.7 Power circuit and voltage feedback

Two 100k ohm resistors in series form a voltage divider that reduces the input voltage by half A 10nF capacitor is added to filter out high-frequency noise, ensuring a stable and clean voltage signal for the ESP32 microcontroller.

2 LDO (Low-Dropout Regulator): For stable power supply

Figure 3.8 Low drop output regulator

Component Assembly

The article highlights the motherboard design, featuring both simulated and assembled images, with final dimensions of 41.25 mm x 66.2 mm that meet project specifications A sequential assembly process is advised, starting with the power block, then the LDO, MCU, sensors, and concluding with the push buttons It is crucial to conduct thorough functionality checks after each assembly stage to ensure optimal performance.

Figure 3.20 Motherboard Design and after Fabrication and Assembly

Figure 3.21 Prototype Hardware Assembly for Animal Wearable System

To ensure accurate and reliable data collection, the following sensor configurations were established:

- MPU6050 and MPU9255 (Accelerometer and Gyroscope): Configured for a sampling frequency of 100 Hz to capture detailed motion data

- NTC Temperature Sensor: Set to a sampling frequency of 1 Hz, suitable for capturing slower-changing temperature variations

- GPS Module: Also configured for a 1 Hz sampling rate to periodically acquire location information

Figure 3.22 Raw data from device

After assembling the motherboard and verifying its functionality, evaluating the quality of the signals it generates is essential A critical factor in this assessment is the signal-to-noise ratio (SNR), which measures the strength of the desired signal relative to the unwanted background noise The SNR can be calculated using a specific formula expressed in decibels (dB).

Where: S is the average value of the signal over a set of samples

N is the noise (A measure of how much the signal fluctuates around the mean)

After calculating for 1000 samples of mpu6050, mpu9255, 50 samples of NTC temperature sensor we have table:

Table 3.3 Sensor Data Quality Analysis: Signal-to-Noise Ratio Measurements

Sensor Data SNR Average SNR

Accelerometer z-axis 46.22 dB Gyroscope x-axis 38.09 dB

Accelerometer z-axis 45.88 dB Gyroscope x-axis 20.52 dB

NTC temperature Temperature 42.04 dB 42.04 dB

Overall, the motherboard demonstrates good SNR performance for most sensors, indicating strong signal quality and minimal noise interference.

ALGORITHMS

Data Processing

4.1.1 MPU6050 Data Filtering and Noise Reduction

Raw sensor data from the MPU6050 can be affected by noise, which may arise from electrical interference, mechanical vibrations, and sensor imperfections, ultimately compromising measurement accuracy To obtain reliable information from this data, a multi-stage filtering and noise reduction process is implemented This process involves a series of filtering steps designed to minimize noise and improve the quality of the sensor data A block diagram illustrates the overall noise reduction process effectively.

Figure 4.2 Data filtering block diagram

The MPU6050 sensor is prone to offset errors when at rest, which can lead to inaccurate readings To ensure precision, calibrating the MPU6050 is essential, as it helps identify and eliminate these unwanted values Gyroscope calibration involves averaging 500 stationary readings and subtracting this average from subsequent readings to correct for bias For accelerometer calibration, 500 samples are collected per axis in three orientations: upward, downward, and perpendicular A linear fit is then applied to the data, allowing for the computation of calibrated acceleration values.

Determination of dog’s movement frequency range: A data-driven approach

To optimize our lowpass filter design, understanding the frequency range of dog movements was crucial We utilized a dataset from a study by the University of Helsinki's Faculty of Veterinary Medicine, which recorded dog movements at a sampling rate of 100 Hz, including acceleration and gyroscope data We prepared the dataset by eliminating any missing or undefined data and focused on the timestamps and three-dimensional acceleration measurements To analyze the various frequencies in the dog's movement, we applied the Discrete Fourier Transform (DFT).

This equation transforms the time-domain signal x(n) into the frequency-domain representation X(k) Each X(k) represents the amplitude and phase of a particular frequency component in the original signal

Frequency resolution (Δf) defines our ability to differentiate between various frequencies and is influenced by the sampling rate (fs = 100 Hz in this instance) and the number of samples (N) To achieve improved frequency resolution, it is essential to either increase the sampling rate or utilize a greater number of samples.

According to the Nyquist theorem, the frequencies we can analyze are restricted to half of the sampling rate, allowing us to accurately detect frequencies up to 50 Hz This limitation helps prevent aliasing, a phenomenon where high-frequency signals are incorrectly identified as lower frequencies.

To visualize the distribution of power (energy) across different frequencies, we calculated the power spectrum (P(k)) This helps us identify which frequencies are most dominant in the dog's movement

We then set a threshold to differentiate meaningful frequencies from background noise Any frequency with a power level above this threshold was considered significant

We established the dog's movement frequency range by identifying the minimum and maximum significant frequencies This range includes the key frequencies associated with the dog's body movements during different activities Our analysis revealed that the highest frequency for rapid movements, like jumping, reached a maximum significant frequency of 18.6 Hz.

Figure 4.4 Investigating the Vibrational Frequencies of Jumping in Dogs: Gyroscope in

A Finite Impulse Response (FIR) filter was utilized to improve the quality of raw MPU6050 sensor data by effectively reducing high-frequency noise This filtering process ensures that crucial low-frequency motion signals, vital for analyzing dog behavior, are preserved.

We utilized the Kaiser window method for designing our FIR low-pass filter, striking an optimal balance between filter performance and computational complexity The design process involves specifying three critical parameters: passband ripple (𝑎 𝑃), stopband attenuation (As), and the transition band Our filter targets a cutoff frequency of 𝐹 𝑐 Hz, with a sampling frequency of 𝐹 𝑠 = 100Hz, a passband from 0 to 20Hz to encompass the dog's expected activity range, a stopband at 30Hz, a passband ripple of ±1dB, and a stopband attenuation of 40dB These carefully selected parameters aim to isolate the frequency range most pertinent to canine behavior while effectively minimizing higher-frequency noise.

𝛥𝜔 = 𝜔 𝑠 − 𝜔 𝑝 = 0.2π (4.8) Convert the passband ripple to linear:

For FIR filters, we can use the Kaiser window method, which provides a good trade-off between filter performance and complexity The formula for estimating the filter order is:

Where: n is the filter length

𝐴 𝑠 is the stopband attenuation in dB

𝛥𝜔 is the normalized transition bandwidth

Round up to the next odd integer: N = 25 so filter order = N - 1 = 24

The Kaiser window parameter β is calculated based on the stopband attenuation:

For our case with 𝐴 𝑠 = 40 dB: 𝛽 ≈ 3.395

The Kaiser window is defined as

The filter coefficients are calculated with the Kaiser window The general formula is:

Where: h[n] are the filter coefficients w[n] is the Kaiser window

ℎ 𝐷 [𝑛] are the ideal impulse responses

Figure 4.5 FIR Filter Frequence response

Figure 4.6 MPU6050 Data Filtering with FIR Low-Pass Filter

The Analog-to-Digital Converter (ADC) in the ESP32 microcontroller is essential for precise sensor readings in pet monitoring systems, yet it exhibits non-linear characteristics that can cause measurement inaccuracies To mitigate this, we have developed a thorough ADC calibration procedure that starts with generating a precise reference voltage using the onboard DAC By recording and averaging multiple ADC readings for each voltage level, we apply linear interpolation to create a continuous calibration curve This process results in a high-resolution Lookup Table (LUT) that enhances sub-LSB accuracy An inverse mapping algorithm identifies the closest actual voltage value in the LUT for each ADC reading, leading to an optimized calibration LUT that balances accuracy and memory usage Ultimately, this calibrated ADC yields significantly more accurate voltage measurements compared to its uncalibrated counterpart.

The calibration results are depicted in the figure below, where the red line indicates the DAC output voltage on IO25, the blue line represents the raw reading from the ESP32, and the green line illustrates the calibrated value.

Figure 4.7 ESP32 ADC Calibration Results

After calibration, the Look-Up Table (LUT) facilitates the conversion of raw ADC values into accurate voltage readings This calibrated voltage is essential for calculating the thermistor resistance (R) using the appropriate formula.

Where: 𝑎ⅆ𝑐 𝑚𝑎𝑥 is the maximum ADC value (4095),

𝑎ⅆ𝑐 𝑣𝑎𝑙𝑢𝑒 is the ADC value after calibrated

𝑅 0 is the thermistor resistance at a reference temperature (𝑇 0 )

The temperature (T) in Kelvin can then be accurately determined using the Beta parameter equation:

The thermistor

Figure 4.8: Application of Averaging Filter to Temperature Data

Machine Learning Model Development for Behavior Prediction

The process of developing a machine learning model for predicting dog behavior involves systematically transforming raw sensor data into actionable insights This section details the steps from data collection to model performance comparison, utilizing SVM and LSTM algorithms The aim is to outline data processing and model training procedures to identify the most effective algorithm for accurate dog behavior prediction Key stages include meticulous data handling, preprocessing, feature extraction, and the application of class weight techniques, all of which contribute to the development and evaluation of predictive models The workflow is visually represented in Figure 4.8, highlighting the journey from raw data to the selection of the best-performing model.

Figure 4.9 Flowchart of Machine Learning Model Development for Behavior Prediction

High-quality data is essential for any AI project, as it drives algorithm development and guarantees the accuracy and reliability of model outcomes Due to constraints in our data collection abilities, we leveraged a dataset from an experimental research study conducted by the University of Helsinki's Faculty of Veterinary Medicine in Finland.

The dataset comprises measurements from gyroscope and accelerometer sensors, collected at a consistent sampling rate of 100 Hz These sensors were strategically placed on the dogs' back and neck to ensure uniform orientation during testing Video recordings were utilized for precise annotation of the dogs' behaviors, resulting in a comprehensive dataset covering 106,110 seconds (approximately 29.48 hours) The annotated data includes various behaviors such as sitting, standing, lying on the chest, walking, trotting, galloping, and sniffing, which will be used to develop machine learning models for classification.

The data was saved in the DogMoveData.csv file This file contains the following columns:

Table 4.1 Descriptions of the columns of DogMoveData.csv file

DogID Number ID of the dog

TestNum Number of the test {1, 2} t_sec Time from the start of the test in seconds

ABack_x Accelerometer measurement from the sensor in the back, x-axis

ABack_y Accelerometer measurement from the sensor in the back, y-axis

ABack_z Accelerometer measurement from the sensor in the back, z-axis

ANeck_x Accelerometer measurement from the sensor in the neck, x-axis

ANeck_y Accelerometer measurement from the sensor in the neck, y-axis

ANeck_z Accelerometer measurement from the sensor in the neck, z-axis

GBack_x Gyroscope measurement from the sensor in the back, x-axis

GBack_y Gyroscope measurement from the sensor in the back, y-axis

GBack_z Gyroscope measurement from the sensor in the back, z-axis

GNeck_x Gyroscope measurement from the sensor in the neck, x-axis

GNeck_y Gyroscope measurement from the sensor in the neck, y-axis

The GNeck_z gyroscope measures sensor data from the neck along the z-axis, capturing specific tasks as they occur When no tasks are being performed, the system remains undefined It allows for the annotation of up to three simultaneous behaviors, categorized as behavior_1, behavior_2, and behavior_3 Additionally, PointEvents are recorded for short, distinct occurrences, such as a bark.

Effective data processing is essential for the success of any AI project, as the quality and organization of data significantly influence model performance and accuracy Proper preprocessing transforms raw data into a structured and clean format, making it suitable for analysis In this project, we focused on three key actions to enhance data quality.

To ensure the quality and relevance of our dataset for developing accurate machine-learning models, we performed several data-cleaning steps:

1 Remove Irrelevant Behavior Annotations: The dataset includes various behaviors, but our study focuses on seven specific behaviors: sitting, standing, lying on the chest, walking, trotting, galloping, and sniffing Other annotated behaviors have been removed to streamline the dataset for our analysis

2 Remove Simultaneous Column Behaviors: Situations where two or three behaviors occur simultaneously have been annotated accordingly in the dataset However, for our analysis, we only need single behavior annotations Therefore, columns representing simultaneous behaviors (behavior_2, behavior_3) have been removed

3 Remove Undefined Annotations: Approximately 34.4% (36,494 seconds) of the data is marked as undefined, indicating no specific behavior annotation This data cannot be used in developing supervised classification algorithms as it may include targeted behaviors without labeling Thus, all rows with undefined annotations have been removed

4 Drop Unnecessary Columns: To further refine the dataset, we removed columns that are not essential for our analysis These columns include: o DogID o TestNum o t_sec o task o behavior_2 o behavior_3 o PointEvent

After these cleaning steps, our dataset is focused solely on the relevant data, ensuring that it is well-prepared for training and developing our machine learning models

Identifying and managing outliers is crucial for maintaining the integrity of data analysis and ensuring model robustness This involves detecting anomalous data points that significantly differ from the rest of the dataset and taking appropriate actions, such as removal or correction, to prevent skewed results.

To demonstrate the effect of outliers, we start with a visual plot of the original data, which aids in comprehending the data distribution and pinpointing any extreme values that could distort the analysis.

Figure 4.11 Plots of Original Back Sensor Data

In this project, we utilize the Z-Score Filtering technique to effectively manage outliers in our dataset This method standardizes the data, allowing us to assess how far each data point deviates from the mean in terms of standard deviations By establishing a Z-score threshold of 3, we can pinpoint and eliminate data points identified as outliers Subsequently, we generate a histogram plot of the filtered data to visually compare it with the original dataset, highlighting the impact of the Z-Score Filtering process.

Figure 4.12 Plots of Back Sensor Data After Applying Z-score filtered Data

Z-score filtering yields positive outcomes by closely aligning with sensor data, preserving the dataset's authentic characteristics while efficiently handling outliers.

After processing, the time series data produced by the movement sensors were saved and analyzed in real-time with Python version 13.12

Feature extraction plays a vital role in machine learning and data analysis, particularly for time series data such as movement sensor signals This process is essential for enhancing model performance, reducing dimensionality, and improving the interpretability of the data before training.

Dimensionality reduction is essential for optimizing raw sensor data, which typically includes numerous data points By extracting key features, we can significantly decrease the data's dimensionality while retaining crucial information This approach enhances the efficiency of the training process and minimizes the risk of overfitting.

To effectively distinguish between different behaviors, it is crucial to capture relevant characteristics of movement, such as total activity, position offset, and mean crossings These higher-level features provide more informative insights compared to raw data points, enhancing the analysis of behavioral patterns.

The time series signals were divided into two-second bins with 50% overlap A total of

Fuzzy Logic for Health Prediction

Numerous studies have shown a significant relationship between a dog's body temperature, behavior, and overall health The normal body temperature for dogs ranges from 37.5°C to 39.5°C, as noted by Greer et al (2007) This temperature can vary with the dog's level of activity, with Piccione et al (2011) finding that physical exertion can raise a dog's body temperature by as much as 2.3°C above its baseline Additionally, Rizzo et al (2017) found that abrupt changes in body temperature and behavior may signal health concerns, including infections, inflammation, or endocrine disorders.

We propose a fuzzy logic system to evaluate a dog's health by integrating 24-hour calorie expenditure data, body temperature, and activity levels Utilizing established metabolic equivalent of task (MET) values, the system categorizes activity intensity as low, moderate, or high, while calculating caloric expenditure based on MET values, weight, time, and resting energy requirements (RER) Additionally, it considers general daily caloric intake guidelines relative to activity levels This comprehensive fuzzy logic model enhances the assessment of a dog's health status, aiding in the early detection of potential health issues and enabling personalized interventions.

We utilized a fuzzy logic approach to model the intricate relationship between environmental temperature, canine behavior, and energy expenditure, effectively addressing the uncertainties and imprecisions present in biological systems This process included the fuzzification of three key input variables through the definition of membership functions.

- Behavior: Characterized by 7 activities are lying, sitting, standing, sniffing, walking, trotting, and galloping

- Energy Expended: Quantified by referencing the Daily Energy Requirement (DER) formula, which utilizes the dog's body weight, activity level, and time spent on each behavior

- Body Temperature: Represented as a range of values reflecting normal dog’s thermoregulation

- An additional output variable, representing the dog's overall health status, was also defined

The behavior membership function utilized Metabolic Equivalent (MET) values, inspired by the research of Gerth et al on Inuit sled dogs We classified behaviors into three distinct intensity levels.

Figure 4.16 Dog’s Metabolic Equivalent of Task

We implemented these categories using Gaussian membership functions:

- Low Intensity: Gaussian function centered at 0 with a standard deviation of 0.6

- Moderate Intensity: Double Gaussian function with means at 2.5 and 3.5, and standard deviations of 0.4 and 0.5 respectively

- High Intensity: Double Gaussian function with means at 6.5 and 8, and standard deviations of 0.6 and 1 respectively

To estimate the energy expended by the dog, we employed the following formulaError! R eference source not found.Error! Reference source not found :

𝐷𝐸𝑅 = 𝑅𝐸𝑅 ∗ 𝐴𝑐𝑡𝑖𝑣𝑖𝑡𝑦 𝑓𝑎𝑐𝑡𝑜𝑟 (4.21) Where: 𝐷𝐸𝑅: Daily Energy Requirement (kcal/day)

RER (Resting Energy Requirement) = 70 ∗ 𝑚 0.75 (kcal/day)

𝑚 is body weight in kg

And the Activity factor varies based on the dog's lifestyle:

1.2-1.4 for inactive/obese prone dogs 1.6-1.8 for moderately active dogs 2.0-5.0 for active working dogs

The total energy expended throughout the day was then calculated by summing the energy consumed during each behavior, weighted by the time spent on that behavior:

Where: 𝛴𝐸: Total Energy Expended (kcal/day)

RER = 70 ∗ 𝑚 0.75 (𝑚 is body weight in kg)

𝑀𝐸𝑇: Metabolic equivalent of task for each behavior

𝑡: time of each behavior (hour)

This comprehensive approach allowed us to model the complex interactions between environmental temperature, canine behavior, and energy expenditure, providing valuable insights into canine physiology and well-being

We implemented these categories using a combination of Gaussian and double Gaussian membership functions:

- Inactive: Double Gaussian function with parameters adjusted based on the inactive energy expended level

- Moderately Active: Gaussian function centered at the average of all three activity levels

- Active: Double Gaussian function with parameters adjusted based on the active energy expended level

The temperature membership function was defined using three categories:

- Low: Double Gaussian function with inflection points at 25°C, 36°C, and 36.5°C

- Normal: Gaussian function centered at 38.5°C with a standard deviation of 0.5

- High: Double Gaussian function with inflection points at 40.5°C, 41.5°C, and 45°C

By encompassing the expected range of measurable temperatures, this method safeguards against code failures and preserves the precision of the defined membership functions

Health Status (Output) Membership Function:

The output variable, health status, was divided into four levels:

- Critical: A Gaussian function centered at 0 with a standard deviation of 1.5 This function represents the most severe health status

- Concerning: A Gaussian function centered at 3.5 with a standard deviation of 1 This function represents a health status that requires attention but is not as severe as critical

- Good: A Gaussian function centered at 6.5 with a standard deviation of 1 This function represents a generally healthy status

- Excellent: A Gaussian function centered at 10 with a standard deviation of 0.5 This function represents the optimal health status

Figure 4.17 Membership Functions Fuzzy Rules

Our fuzzy logic system is built on 27 meticulously crafted rules that establish the connections between input variables—behavior intensity, energy expenditure, and body temperature—and the output variable of health status These rules are designed to accurately reflect the intricate interactions between a dog’s physiology and environmental influences, capturing various physiological states and their significance for canine health.

- In “Low” temperature conditions, “High Intensity” behavior combined with “Inactive” energy expenditure results in a “Critical” health status, reflecting the risk of hypothermia and energy depletion

- Under “Normal” temperature conditions, “Moderate Intensity” behavior with “Active” energy expenditure leads to an “Excellent” health status, indicating an ideal balance of activity and nutrition

- In “High” temperature scenarios, “Low Intensity” behavior with “Inactive” energy expenditure results in a Critical health status, highlighting the risk of heat stress and metabolic imbalance

These guidelines provide a detailed framework for assessing a dog's health under various environmental and physiological conditions By utilizing fuzzy logic, these rules enable a nuanced understanding, allowing for the simultaneous consideration of partial membership across multiple categories.

The system conducts a daily assessment of the dog's health by analyzing calorie intake over a 24-hour period, along with temperature and behavioral data The evaluation categorizes the dog's health into one of four levels: Critical, Concerning, Good, or Excellent.

This method enables a thorough evaluation of a dog's health by monitoring both body temperature and activity levels over an entire day, facilitating early identification of potential health concerns and prompt notifications to pet owners For a dog weighing 12kg, we have obtained the following results:

Figure 4.19 Health Status Prediction Results Using Fuzzy Logic

DEPLOYMENT ON AWS

System Architecture Overview

Figure 5.1 Integrated Architecture for Pet Monitoring

The advanced smart wearable device for dog monitoring utilizes a complex connection flow to relay data from the device to the cloud and then to the user interface Central to this system is the wearable device, which features sensors like accelerometers, GPS, and temperature sensors that gather crucial information regarding the dog's activity, location, and health Initially, this data is processed by the device's ESP32 microcontroller before being sent to the cloud through Wi-Fi or NB-IoT, depending on the device's setup.

The data journey begins with a secure connection to AWS IoT Core, utilizing the MQTT protocol for transmission Upon arrival, predefined IoT rules dictate the data's processing and routing, triggering AWS Lambda functions for tasks like validation and transformation The processed data is stored in Amazon DynamoDB, ensuring scalable NoSQL database solutions for real-time and historical data For advanced data analysis, including machine learning model implementation, an EC2 instance is used Communication between backend services and user-facing applications is facilitated by Amazon API Gateway, which provides RESTful API endpoints for data retrieval and device management.

Web and mobile applications utilize API endpoints to retrieve processed data, delivering real-time updates, historical trends, and alerts to users in a user-friendly manner This robust architecture guarantees an efficient and secure data flow from wearable devices to end users, employing AWS services to create a scalable and reliable system for real-time dog monitoring.

Device Configuration and Connectivity

We created a custom device configuration interface that enables users to easily set up and tailor wearable devices to their specific needs and network environments This interface features Wi-Fi and NB-IoT mode selection, network credential input, and an advanced developer mode Activation occurs when a physical button on the device is pressed, prompting the ESP32 to generate a Wi-Fi access point Users can connect to this network and access a web-based configuration page for seamless customization.

Our pet monitoring device features an intuitive setup process that allows users to easily select between Wi-Fi and NB-IoT connectivity options, input network credentials, and access advanced settings if needed With intelligent network switching, the device automatically transitions to a 4G network if the Wi-Fi connection is lost for over one minute, ensuring continuous data transmission This seamless connectivity enhances the reliability of our system, providing pet owners with uninterrupted updates even in areas with unstable Wi-Fi The user-friendly configuration interface allows for quick setup and customization, ultimately improving the overall user experience of our pet monitoring solution.

Figure 5.2 Push the button to configure the device

• User initiates configuration by pressing the physical button on the device

• ESP32 creates a Wi-Fi access point

• User connects to the access point using a smartphone or computer

• A web-based configuration page automatically loads

• User selects the operating mode (Wi-Fi or NB-IoT) and enters necessary details

• For Wi-Fi mode, users can choose between AWS (for end-users) or Mosquito (for developers) sub-options

• Device saves the configuration and restarts in the selected mode

Figure 5.3.Configuration Interface for the Dog Monitoring Device

Cloud-based Data Processing and Storage

Figure 5.4 AWS-based IoT Architecture for Smart Wearable Devices

Our deployment and cloud services system aims to establish a secure, scalable platform for collecting, processing, and analyzing data from smart wearable devices for dogs This system guarantees smooth data transmission to users, enabling real-time analysis and long-term data storage.

5.3.1 Data Ingestion and Deployment (AWS IoT Core)

AWS IoT Core is a managed cloud service that facilitates secure, bi-directional communication between Internet-connected devices and the AWS Cloud, acting as the gateway for data from our pet monitoring system's wearable device into the AWS ecosystem.

Configuring AWS IoT Core aims to create a secure and reliable connection between wearable pet monitoring devices and AWS cloud infrastructure, facilitating real-time data ingestion, device management, and integration with other AWS services The system processes raw data from these devices, including GPS coordinates, accelerometer data, temperature, and device status information like battery level The output consists of authenticated, standardized data formats and triggers for AWS Lambda functions to enable further processing.

• Generate unique X.509 certificates and private key pairs for each wearable pet monitoring device [29]

• Securely store these credentials on the devices during the manufacturing or initialization process

• Register each device in AWS IoT Core, creating a distinct "thing" with its associated certificate

• Ensure the device firmware is configured to use these credentials for authentication with IoT Core

• Create an IoT policy that defines the precise permissions for the pet monitoring devices, including:

1 Authorization to connect to IoT Core

2 Permission to publish data to specific MQTT topics

3 Ability to subscribe to necessary topics

• Attach this policy to each device's certificate to enforce access control

• Set up additional policies for AWS services (e.g., Lambda, DynamoDB) that will interact with IoT Core

• Design an MQTT topic structure tailored to the pet monitoring system

• Implement IoT Rules to process incoming MQTT messages

• Configure error actions for each rule to handle message processing failures and ensure data integrity

• Configure the device firmware to use MQTT over TLS for secure communication with IoT Core

• Implement X.509 certificate-based authentication in the device firmware

• Set up the devices to use port 8883 for MQTT communication, which is the standard port for MQTT over TLS

• Ensure the device firmware validates the AWS IoT Core server certificate to prevent man-in-the-middle attacks

• Conduct thorough testing of the device-to-cloud communication, verifying successful authentication and data transmission

• Simulate various scenarios, including message floods and network interruptions, to ensure robust performance

• Validate that all IoT Rules are functioning correctly and routing data to the appropriate AWS services

Figure 5.5 Chart showing AWS IoT Core performance metrics

The implementation of IoT Core configuration achieved remarkable results, particularly in secure device connectivity and data ingestion performance An impressive 94% of registered devices established encrypted MQTT connections to AWS IoT Core, with zero unauthorized access attempts during initial testing The system showcased strong data handling capabilities, processing an average of 10 messages per second from devices, while maintaining a message latency of just 150ms under normal conditions These results highlight the effectiveness of the data ingestion and deployment layer in reliably and efficiently collecting data from multiple wearable devices simultaneously.

5.3.2 AWS Lambda Functions for Data Processing

AWS Lambda is a vital component of our pet monitoring system, enabling code execution without the need for server management This serverless computing service efficiently processes data from IoT devices and seamlessly integrates with various AWS services.

The implementation of Lambda functions in our system aims to efficiently process and transform sensor data from wearable devices, ensure seamless integration between various AWS services, and provide a scalable backend for data processing and API requests These functions utilize raw sensor data from IoT Core for data ingestion and API requests from API Gateway for integration, producing processed data stored in DynamoDB and formatted responses for API requests.

• Create a Lambda function triggered by IoT Core rules

• Implement data validation and basic transformation logic

• Configure DynamoDB integration using AWS SDK

• Set up error handling and retry mechanisms

• Create a Lambda function to handle API requests

• Implement authentication and authorization checks

• Develop data retrieval logic from DynamoDB

• Implement response formatting and error handling

• Set up execution roles and permissions for accessing AWS services

• Configure environment variables for sensitive information

• Set appropriate memory and timeout settings

• Enable AWS X-Ray for distributed tracing

• Set up CloudWatch Logs for each Lambda function

• Configure CloudWatch Alarms for error rates and duration thresholds

• Implement custom logging for important events and errors

• Minimize external dependencies in Lambda functions

• Implement efficient coding practices (e.g., connection reuse)

• Consider using Provisioned Concurrency for consistent performance

The implementation of AWS Lambda functions in our pet monitoring system significantly enhanced data processing efficiency and API response performance, resulting in faster processing times and a more seamless user experience Error rates decreased notably, with most function invocations completed successfully, while overall system costs were reduced compared to traditional server-based solutions This serverless approach has provided our pet monitoring system with a scalable, efficient, and cost-effective solution for data processing and API integration.

Amazon DynamoDB is a fully managed NoSQL database service that delivers fast and predictable performance with seamless scalability It is essential for our pet monitoring system, as it efficiently stores and retrieves sensor data and analysis results.

The implementation of DynamoDB tables in our system aims to achieve real-time storage of raw sensor data from wearable devices, facilitate the storage of behavior analysis results from machine learning models, and enable efficient data querying and retrieval for analysis and visualization By utilizing Lambda functions, the system processes raw sensor data and behavior analysis results, ultimately providing stored data and query results for client applications and analytical tools, all while ensuring scalability and high performance in data operations.

• Create a table with a flexible schema to accommodate various sensor data types

• Define the primary key as a composite key: (timestamp as sort key)

• Include attributes for accelerometer, gyroscope, GPS, temperature, and battery level data

• Create a table to store machine-learning model outputs

• Define the primary key as a composite key: (timestamp as sort key)

• Include attributes for behavior classification and health status assessment

• Set up appropriate read and write capacity units based on expected data volume

• Enable DynamoDB auto-scaling to handle varying workloads

• Implement Time-to-Live (TTL) for data retention management

• Develop Lambda functions to process and store incoming sensor data

• Implement batch write operations for efficient data insertion

• Ensure data consistency and handle potential write conflicts

• Create global secondary indexes for common query patterns

• Implement efficient querying mechanisms for time-based data retrieval

Figure 5.6 Sample Sensor Data On DynamoDB

The integration of Amazon DynamoDB tables into our pet monitoring system has led to remarkable performance enhancements We achieved an impressive 94% efficiency in data ingestion, with write operations completing in under 10ms Query performance also improved significantly, with an average latency of just 5ms for accessing recent data within the last 24 hours Historical data queries, covering up to 15 days, were consistently completed in under 100ms, allowing for quick access to long-term trends Data consistency was exceptional, boasting a 98% consistency rate for sensor readings, and during a rigorous 15-day continuous operation test, we experienced zero data loss, highlighting the reliability of our storage solution.

5.3.4 API Gateway for Data Access

Amazon API Gateway is a fully managed service that simplifies the creation, publication, maintenance, monitoring, and security of APIs at any scale It acts as the entry point for applications to access essential data, business logic, and functionalities from backend services within our pet monitoring system.

The implementation of an API Gateway aims to create a secure and scalable interface for client applications to access pet monitoring data, while effectively managing backend service access, API versioning, throttling, and caching to enhance performance and reliability This system processes HTTP/HTTPS requests from both web and mobile client applications, as well as responses from backend integrations like Lambda functions and DynamoDB The output consists of RESTful API responses for client applications, along with metrics and logs that track API usage and performance.

• Define API resources and methods (GET, POST, PUT, DELETE) for pet monitoring data access

• Create API models to define the structure of request/response payloads

• Set up API stages for development, testing, and production environments

• Configure Lambda function integrations for data processing and retrieval

• Set up direct integrations with DynamoDB for simple read operations

• Implement request/response mapping templates for data transformation

• Configure API keys for client application authentication

• Implement AWS IAM roles and policies for fine-grained access control

• Set up AWS Cognito integration for user authentication and authorization

The integration of Amazon API Gateway into our pet monitoring system has significantly improved API performance, resulting in reduced latency and enhanced request processing efficiency, which contributes to a smoother user experience Enhanced security measures effectively prevent unauthorized access, reinforcing the system's integrity Additionally, the developer experience has greatly improved, with positive feedback on API usability and documentation, streamlining maintenance and enabling future expansions and integrations.

5.3.5 Model Deployment on AWS EC2

Amazon Elastic Compute Cloud (EC2) powers our pet monitoring system by hosting and executing the machine learning model for behavior prediction, offering dedicated computational resources and flexibility for optimal model performance.

Deploying our behavior prediction model on AWS EC2 aims to create a dedicated environment for real-time sensor data processing and behavior analysis By utilizing Lambda functions to gather data from wearable devices and leveraging a pre-trained machine-learning model, we can generate accurate behavior predictions EC2's flexibility and scalability enable us to adjust computational resources dynamically, ensuring both optimal performance and cost-efficiency for our model execution in a secure and controlled setting.

• Choose an appropriate EC2 instance type based on model requirements

• Set up the instance with a stable operating system (e.g., Amazon Linux 2 or Ubuntu Server LTS)

• Configure security groups to restrict inbound traffic to necessary ports and services

• Securely transfer the pre-trained behavior prediction model to the EC2 instance

• Install required dependencies and libraries for model execution

• Set up a designated directory structure for model storage and execution

• Develop a Python script to handle the prediction process

• Implement efficient data loading, model inference, and result formatting

• Optimize the script for real-time processing of incoming requests

• Create API routes for receiving sensor data and returning predictions

• Set up error handling and logging for the API

• Configure IAM roles and policies for EC2 instance access

• Implement encryption for data in transit using HTTPS

• Set up VPC endpoints for secure communication with other AWS services

• Implement auto-start scripts for the prediction service on instance boot

• Set up monitoring and auto-recovery mechanisms for the EC2 instance

• Implement logging and alerting for service status and performance

Figure 5.7 EC2 Model Deployment Performance Scores

The deployment of our behavior prediction model on AWS EC2 led to several positive outcomes:

• Prediction performance improved, with faster processing times for incoming requests

• The system demonstrated good scalability, maintaining performance under increased prediction loads

• High availability was achieved, with minimal downtime over extended periods

• Security measures proved effective, with no unauthorized access recorded.

User Interface Development

Our pet monitoring system features two user-friendly interfaces: a comprehensive web application and a streamlined mobile application for Android devices The web application is accessible from any device with a web browser, providing full functionality and detailed visualizations In contrast, the mobile app offers real-time alerts and allows users to monitor their pets' activities on the go Both interfaces prioritize user experience with intuitive navigation, clear data presentation, and responsive design to suit various devices and screen sizes.

The web application enables users to monitor their dog's health and behavior in real time, access historical data, and manage the device seamlessly Designed for accessibility from any web browser, it ensures a consistent experience across various devices and platforms.

• Displays real-time data including health status, current behavior, and body temperature

• Features an interactive map showing the dog's location

• Shows the battery status of the wearable device

• Displays notifications about device temperature, health, and battery status abnormalities b) Charts and Analysis Page:

• Presents temperature trends over time using line charts

• Displays behavior analysis through donut charts

• Offers a color-coded calendar for health status history (can select the date to view temperature graphs and behavior statistics for a specific day)

• Shows a weekly calorie chart for the dog, allowing users to track their pet's energy intake over time

Our back-end system leverages multiple AWS services to establish a robust and scalable architecture, utilizing AWS IoT Core for real-time data streaming, AWS Lambda for serverless computing, Amazon DynamoDB for efficient data storage and retrieval, and Amazon API Gateway for RESTful API endpoints This powerful combination ensures a reliable and responsive back-end infrastructure for our livestock tracking system The web application plays a vital role in offering users a comprehensive overview of their pets' health, behavior, and nutrition, and its accessibility from any web browser allows for monitoring across various devices, making it an essential tool for pet care.

Figure 5.8.User Interface Design for the Pet Monitoring Web Application

The mobile app provides the accessibility of the monitoring system, allowing users to stay connected with their pets' health and location even when away from a computer

The mobile application is created by transforming the web application into a native Android app through VoltBuilder, enabling quick development while ensuring consistency with the web interface This process is illustrated in Figure 54 below.

Figure 5.9 The process of building the mobile app User Interface Optimization:

Special attention is given to optimizing the UI for touch interactions and smaller screens, ensuring a seamless user experience on mobile devices

Figure 5.10 User Interface (UI) of Mobile Application

EXPERIMENTING RESULTS

Wearable Hardware Device Performance

This section evaluates the performance of our developed wearable device for dog monitoring

We conducted a series of tests to assess the accuracy, reliability, and efficiency of the integrated sensors and overall device functionality

• Accelerometer: We tested the accelerometer's accuracy in detecting various dog activities (walking, running, resting) across 100 trials The sensor demonstrated 95% accuracy in distinguishing between these basic activities

GPS location accuracy was evaluated in both open and urban environments, revealing an average deviation of 2.3 meters in open areas and 4.7 meters in urban settings These deviations fall within acceptable ranges for effective pet tracking.

• Temperature Sensor: Body temperature measurements were compared against a veterinary-grade thermometer Our sensor showed an average deviation of ±0.3°C across 50 measurements, which is suitable for detecting significant temperature changes

Battery Life and Power Consumption: The device achieved an average battery life of 4 hours under normal usage conditions, which included periodic GPS updates and continuous accelerometer monitoring

The device demonstrated impressive durability by passing 20 impact resistance tests from a height of 1.5 meters without losing functionality Additionally, a week-long comfort trial involving three dogs of different breeds revealed no signs of skin irritation or discomfort, highlighting its user-friendly design.

Sensor Data Processing and Feature Extraction

Our data processing algorithms were evaluated for their effectiveness in noise reduction and feature extraction from raw sensor data

Noise Reduction: A low-pass filter for accelerometer data reduced high-frequency noise by 60.6%, improving the signal-to-noise ratio of MPU6050 from 36.155dB to 40.20dB

Our algorithm excels in motion classification by extracting essential features from accelerometer data, achieving an impressive 92% accuracy in identifying five distinct dog behaviors: walking, running, playing, resting, and eating, based on 1000 samples Additionally, the system demonstrates its effectiveness in recognizing location patterns, achieving an 88% accuracy rate in location extraction.

Machine Learning Model for Behavior Classification

We evaluated our machine learning model's performance in classifying dog behaviors and detecting potential health issues

• Behavior Classification Accuracy: Overall Accuracy: The model achieved 89% accuracy in classifying seven different dog behaviors (walking, running, playing, resting, eating, barking, and sleeping)

• Health Issue Detection: The model demonstrated 78% sensitivity and 85% specificity in detecting potential health issues based on unusual behavior patterns and temperature variations.

Cloud Integration and Real-time Monitoring

This section assesses the performance of our AWS-based cloud infrastructure in handling data from the wearable device

Data transmission reliability is significantly impacted by cellular coverage, with a success rate of 98% for data packets in areas with strong signals Conversely, in regions with weak coverage, this rate decreases to 95% However, the system effectively queues and transmits data once connectivity is reestablished.

Real-time Processing and Alerts: The system processed incoming data and generated alerts for unusual behavior or potential health issues with an average delay of 2.3 seconds

Data Storage and Retrieval: Query response times for historical data remained under 500ms for datasets up to 1 year old, meeting our performance targets for the web and mobile applications

User Interface Performance: The web application loaded real-time data updates with an average delay of 1.2 seconds, while the mobile app achieved a slightly better performance at 0.9 seconds

In summary, our experimental findings indicate that the developed system successfully achieves the majority of our initial goals, delivering precise, dependable, and prompt monitoring of canine behavior and health Future enhancements should focus on minimizing false positives in health issue identification and optimizing battery longevity.

CONCLUSION AND FUTURE WORK

Summary and Conclusions

This project has successfully created a smart wearable device for real-time monitoring of dogs' health and behavior Key features include hardware integration, real-time data transmission through Wi-Fi and NB-IoT, and cloud-based analytics utilizing AWS services The user-friendly web and mobile interfaces enhance accessibility, while optimized battery efficiency ensures prolonged use This system marks a significant advancement in pet health technology, offering data-driven insights that enable pet owners and veterinarians to make informed decisions regarding dog health and well-being.

Limitations and Challenges

While the project achieved its primary objectives, several limitations and challenges were identified:

• Battery Life: Despite optimization efforts, the need for continuous data transmission and GPS tracking still poses challenges for extending battery life beyond 6 hours

• Size and Comfort: Miniaturizing the device while maintaining all functionalities proved challenging, potentially limiting its suitability for very small dog breeds

• Data Accuracy: Environmental factors and individual dog characteristics can sometimes affect sensor readings, leading to potential inaccuracies in behavior classification, particularly in novel environments

The system's functionality is heavily dependent on Wi-Fi or cellular networks, which can be restricted in areas with weak connectivity; however, the NB-IoT module significantly alleviates these limitations.

• Cost Considerations: The use of high-quality components and cloud services results in a product that may be relatively expensive for widespread adoption.

Future Enhancements and Research Directions

Future research and development opportunities include:

• Integrating advanced sensors for more comprehensive health monitoring

• Improving machine learning models for behavior recognition

• Implementing edge computing to reduce cloud dependency

• Advancing battery technology for extended operational time

• Further miniaturization to suit a broader range of dog breeds

• Adapting the technology for other pet species

• Integrating with smart home systems

• Enhancing predictive health analytics capabilities

• Implementing secure social sharing features

In summary, the existing pet health monitoring system represents a major advancement, yet there is considerable opportunity for future innovation Emphasizing the need to overcome current limitations, broaden functionality, and prioritize ethical, user-friendly, and animal-centered technological development will drive progress in this field.

[1] Whistle, "Whistle GO Explore," Whistle Official Website, 2023

[2] FitBark, "FitBark 2 Dog Activity Monitor," FitBark Official Website, 2023

[3] Ollerenshaw, D., Barrett, D., & Robinson, D (2020) Machine Learning Approaches for the Classification of Canine Behaviors Using Accelerometer Data Animals, 10(6),

[4] Barker, D J., Amundin, M., & Lohilahti, H (2019) Comprehensive Monitoring of Dog Health and Activity Using Wearable Technology PLoS ONE, 14(3), e0212870

[5] Nguyen, T.H., Tran, Q.D., & Le, M.T (2023) "Current State and Future Prospects of Pet Wearable Technology in Vietnam." Vietnam National University Ho Chi Minh City - University of Science, Research Report Series, 25(4), 112-125

[6] Grand View Research (2020) Pet Wearable Market Size, Share & Trends Analysis Report By Technology (RFID, GPS, Sensors), By Application (Identification & Tracking, Behavior Monitoring), By Region, And Segment Forecasts, 2020 - 2027

[7] Smart Connected Pet Collar Market size, share, growth, trends and forecast 2031 -

AndyJassi - wattpad (n.d.) Wattpad.com Retrieved June 30, 2024, from https://www.wattpad.com/story/366241218-smart-connected-pet-collar-market-size- share

[8] PetPace, "PetPace Smart Collar," PetPace Official Website, 2023

[9] 6 reasons for Cisco Wi-Fi 6 (2019, August 5)

[10] NEVSemi (n.d.) ESP32 vs STM32 Nevsemi.com Retrieved July 1, 2024, from https://www.nevsemi.com/blog/esp32-vs-stm32

[11] (N.d.-b) Digikey.com Retrieved July 1, 2024, from https://www.digikey.com/en/maker/projects/comparing-microcontrollers-what-brain- should-i-go-with/02d2dcb1a0d441f5a11fc9956559b226

[12] Singh, G (2023, December 19) Types of wireless communication technologies Telecom Trainer https://www.telecomtrainer.com/types-of-wireless-communication- technologies/

[13] Narrow Band (NB) IoT – The next level Communication Network for Internet of

Things (n.d.) Circuitdigest.com Retrieved June 30, 2024, from https://circuitdigest.com/article/narrow-band-nb-iot-next-level-communication- network-for-internet-of-things

[14] Li, C., Wang, J., Wang, S., & Zhang, Y (2024) A review of IoT applications in healthcare Neurocomputing, 565(127017), 127017 https://doi.org/10.1016/j.neucom.2023.127017

[15] Cloud Application Services (n.d.) CodeKing Solutions Retrieved June 30, 2024, from https://www.codekingsolutions.com/cloud-application-services.html

[16] (N.d.) Researchgate.net Retrieved June 30, 2024, from https://www.researchgate.net/publication/283236612_Integration_of_Cloud_computi ng_and_Internet_of_Things_A_survey

[17] Chen, X., Cao, Y., Pu, J., & Wang, X (2020) Classifying dairy cow behaviors using supervised learning algorithms based on accelerometer data Journal of Dairy Science,

[18] Dogan, S., Cifti, E., & Karaboga, D (2019) Classification of dog behaviors using neural networks and wearable sensor data Sensors, 19(3), 734

[19] Martiskainen, P., Jọrvinen, M., & Skửn, J (2018) Using k-means clustering to identify anomalies in cattle movement data Computers and Electronics in

[20] Kim, J., Lee, H., & Park, S (2021) Behavior classification of laboratory mice using convolutional neural networks PLOS ONE, 16(2), e0246550

[21] Rahman, A., Roy, K., & Halder, S (2021) Identifying stress-related vocalizations in pigs using recurrent neural networks Animal, 15(6), 100312

[22] (Support Vector Machine Algorithm, n.d.) https://www.javatpoint.com/machine- learning-support-vector-machine-algorithm

[23] (Lý thuyết về mạng LSTM) https://phamdinhkhanh.github.io/2019/04/22/Ly_thuyet_ve_mang_LSTM.html

[24] Hochreiter, S., & Schmidhuber, J (1997) Long Short-Term Memory Neural Computation, 9(8), 1735-1780

[25] (Stackoverflow ) https://i.sstatic.net/3NlSp.png

[26] (Deep Learning Cơ Bản, 2019) https://nttuan8.com/bai-13-recurrent-neural- network/

[27] moise-castr (n.d.) Waveshare 10 DOF IMU Sensor (C) inertial measurement unit low power consumption eBay Retrieved July 1, 2024, from https://www.ebay.com/itm/261834627241

[28] SIM7080G Cat-M/NB-IoT HAT (EN) NB-IoT / Cat-M(eMTC) / GNSS HAT for

The Raspberry Pi is a versatile and globally applicable platform that supports various IoT applications It can be enhanced with the SIM7080G CAT-M NB-IoT GNSS HAT, which enables efficient connectivity and location tracking This combination is ideal for developers looking to create innovative solutions in the Internet of Things space For more information, visit ProE's website.

[29] Ratasuk, R., Mangalvedhe, N., Zhang, Y., Robert, M., & Koskinen, J.-P (2016) Overview of narrowband IoT in LTE Rel-13 2016 IEEE Conference on Standards for

[30] SIM7080G Module LPWA Wireless Solutions (n.d.) Simcom.com Retrieved June

30, 2024, from https://www.simcom.com/product/SIM7080G.html

[31] Matek gps module Sam-m8q (n.d.) Boulder MultiRotor LLC Retrieved July 1, 2024, from https://bouldermultirotor.com/product/matek-gps-module-sam-m8q

[32] SAM-M8Q module (2016, August 11) U-Blox https://www.u- blox.com/en/product/sam-m8q-module

[33] (N.d.) Espressif.com Retrieved July 3, 2024, from https://www.espressif.com/sites/default/files/documentation/esp32-wroom-

32d_esp32-wroom-32u_datasheet_en.pdf

[34] Santos, S (2021, March 9) ESP32 OTA (over-the-air) updates - AsyncElegantOTA arduino Random Nerd Tutorials https://randomnerdtutorials.com/esp32-ota-over- the-air-arduino/

[35] Lab, M (2019, March 5) Push button with ESP32 - GPIO pins as digital input Microcontrollers Lab https://microcontrollerslab.com/push-button-esp32-gpio- digital-input/

[36] AWS "Device Authentication and Authorization for AWS IoT." Amazon Web Services, Inc., 2023

[37] AWS Lambda Developer Guide https://docs.aws.amazon.com/lambda/latest/dg/welcome.html

[38] Greer, R J., Cohn, L A., Dodam, J R., Wagner-Mann, C C., & Mann, F A

(2007) Comparison of three methods of temperature measurement in hypothermic, euthermic, and hyperthermic dogs Journal of the American Veterinary Medical Association, 230(12), 1841-1848

[39] Piccione, G., Casella, S., Panzera, M., Giannetto, C., & Fazio, F (2012) Effect of moderate treadmill exercise on some physiological parameters in untrained beagle dogs Experimental Animals, 61(5), 511-515

[40] Rizzo, M., Arfuso, F., Alberghina, D., Giudice, E., Gianesella, M., & Piccione, G

(2017) Monitoring changes in body surface temperature associated with treadmill exercise in dogs by use of infrared methodology Journal of Thermal Biology, 69, 64-

[41] Nutrient requirements of dogs and cats (2006) National Academies Press

A study conducted by Wakshlag et al (2012) assessed the dietary energy intake and physical activity levels of dogs participating in a controlled weight-loss program Published in the Journal of the American Veterinary Medical Association, the research highlights the importance of monitoring both nutrition and exercise to effectively manage canine obesity The findings underscore the need for tailored dietary plans and structured physical activities to achieve successful weight loss in dogs For more details, refer to the full article [here](https://doi.org/10.2460/javma.240.4.413).