Received 1 March 2024, accepted 8 March 2024, date of publication 13 March 2024, date of current version 21 March 2024.

Digital Object Identifier 10.1109/ACCESS.2024.3376707

Toward Intelligent Monitoring in IoT:

AI Applications for Real-Time
Analysis and Prediction
WILLIAM VILLEGAS-CH 1 , JOSELIN GARCÍA-ORTIZ 1,

AND SANTIAGO SÁNCHEZ-VITERI2

1 Escuela de Ingeniería en Ciberseguridad, FICA, Universidad de Las Américas, Quito 170125, Ecuador
2 Departamento de Sistemas, Universidad Internacional del Ecuador, Quito 170411, Ecuador
Corresponding author: William Villegas-Ch ([email protected])

ABSTRACT In the contemporary era, the intersection of the Internet of Things and artificial intelligence
revolutionizes how industries monitor and optimize their operations. In this work, we present a system
that combines real-time monitoring provided by Internet of Things devices with predictive analytics based
on artificial intelligence. This system detects anomalies in real-time and anticipates possible failures,
allowing proactive interventions to maximize efficiency and minimize operating costs. Our findings reveal
a significant improvement in the early detection of abnormal trends, as the system consistently identifies
potential problems long before they become critical failures. Our evaluation employed data sets collected
from controlled and industrial production environments, with more than 1 million records, including critical
parameters such as temperature, humidity, and pressure. The results highlight a significant improvement in
the early detection of abnormal trends, with a temperature detection accuracy of 98.7%, exceeding reference
values and demonstrating the system’s effectiveness in preventing critical failures. The analysis also revealed
previously unrecognized operational patterns, offering opportunities for industrial process optimization. This
work highlights the effective integration of the Internet of Things and artificial intelligence to improve
industrial monitoring, highlighting the tangible benefits of such integration, such as the adaptability and
continuous learning of the system, ensuring its long-term effectiveness.

INDEX TERMS Adaptive feedback, real-time precision, intelligent adaptation, IoT-based optimization.

I. INTRODUCTION of technicians and operators. However, this method has

In the current era of digitalization, the automation and limitations, mainly when dealing with complex or large-scale
optimization of processes in various industries have become systems where multiple parameters must be monitored
essential to remain competitive. The adoption of emerg- simultaneously [2]. Furthermore, the evolution of Industry
ing technologies has been a constant in the search to 4.0 and the growing demand for intelligent and connected
improve efficiency, reduce costs, and guarantee quality [1]. solutions have highlighted the need for more advanced
Within this panorama, real-time monitoring of machines and systems. Previous studies have demonstrated the advan-
systems has been positioned as a critical tool, especially tages of implementing monitoring tools based on emerging
in sectors where precision and speed in decision-making technologies such as the Internet of Things (IoT) and
can make the difference between success and failure. Artificial Intelligence (AI) [3]. These tools enable real-time
Historically, equipment and processes have been monitored data collection and offer advanced analysis and prediction
manually, relying heavily on the judgment and experience capabilities [4].
In response to this scenario, this work proposes to
The associate editor coordinating the review of this manuscript and develop an AI-based filtering system specifically designed
approving it for publication was M. Anwar Hossain . for real-time event monitoring. Unlike conventional systems,
2024 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
40368 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ VOLUME 12, 2024
W. Villegas-Ch et al.: Toward Intelligent Monitoring in IoT: AI Applications

this tool continually adapts and learns from the data collected, for future advances in intelligent automation and process
ensuring superior accuracy and efficiency. It integrates optimization.
multiple sensors, including IoT cameras and accelerometers,
to provide a holistic and detailed view of the operation [5]. II. MATERIALS AND METHODS
Although numerous monitoring systems are on the market, The methodology adopted for this study incorporates a
integrating multiple sensors with self-adjustment and learn- combination of advanced tools and techniques. A series
ing capabilities represents an innovative proposal. Therefore, of IoT sensors were used for real-time data collection,
this research provides empirical data on the operation and each meticulously calibrated to ensure accuracy in readings.
effectiveness of such a system and introduces a novel Strategically dispersed devices captured vital parameters like
perspective on how emerging technology can overcome temperature, humidity, and vibration patterns. Data collection
traditional challenges in industrial monitoring and control [6]. was complemented by a filtering system based on artificial
This tool detects and predicts trends and patterns, allowing intelligence, designed to process and analyze the information.
early interventions and avoiding costly downtime or failures. This system incorporates feedback mechanisms and is
Regarding innovation, the combined use of IoT and AI in our constantly adapted to optimize the accuracy and efficiency
proposal marks a new standard in real-time monitoring, posi- of data transmission. The collaboration of technicians and
tioning it as one of the appropriate solutions for the current users was essential, providing crucial feedback that fed and
market [7]. refined the model over time. Implementation and validation
After analyzing the existing literature, it is evident that were carried out in controlled environments, ensuring that any
although many researchers have addressed the challenge findings or patterns identified were based on accurate and
of real-time monitoring, few have effectively integrated replicable situations.
AI adaptability and continuous learning capabilities [8].
Many systems focus on data collection and presenta- A. DEFINITION OF THE PROBLEM
tion but lack our tool’s deep analysis and prediction IoT has revolutionized the technology industry and how we
capabilities. Furthermore, integrating multiple sensors in interact with our environment. IoT devices have infiltrated
our proposal allows a detailed and precise view of various sectors, from smart homes to industrial and health-
the operation, overcoming the limitations of traditional care. However, one of the most pressing challenges is the
solutions [9]. massive volume of data these devices constantly generate
Unlike previous studies that focus on post-processing and transmit. This constant transmission not only consumes
performance or IoT systems operating in isolation, our large amounts of bandwidth but can also incur additional
work coherently integrates real-time AI processing with costs, cause latency, and overwhelm networks, especially
IoT devices. This synergy creates a more robust and when much of the data transmitted may not be immediately
efficient system, overcoming the limitations of conventional relevant.
approaches that did not consider the dynamic interaction Given this, the need arises for systems that can discern
between these two components. between critical data that must be transmitted immediately
This work has demonstrated through practical implemen- and those that are not. Implementing such systems would
tation and constant feedback that the system is theoretically relieve networks from unnecessary data overload and ensure
sound, applicable, and valuable in natural environments. faster and more efficient responses to abnormal or critical
Additionally, significant improvements in the efficiency of events. However, developing these systems is not trivial, as it
real-time transmissions have been identified, with notable requires a combination of application domain understanding,
reductions in interruptions and failures [10]. The system has advanced data processing knowledge, and machine learning
demonstrated increasing adaptability and accuracy through skills.
continuous feedback and learning, validating its effectiveness Given this scenario, the central question arises: How can an
and potential. AI-based adaptive system be developed and implemented in
According to the results, this system has immediate prac- IoT devices that filter and categorize data in real-time before
tical applications, especially in industries where real-time its transmission to optimize bandwidth use and guarantee
monitoring is crucial, such as manufacturing, logistics, and quick responses to abnormal events? To address this issue,
critical infrastructure management. By solving problems this proposal has considered aspects such as selecting the
such as data transmission latency and network overload, most appropriate AI model to operate in real-time on IoT
our approach not only improves operational efficiency but devices and the strategies to efficiently adapt to changing
also reduces costs and improves decision-making based on environmental patterns and feedback.
accurate and timely data. To more clearly visualize how this system can practically
The implications of this study are vast and multifaceted. and directly impact a real scenario, it is essential to consider
It opens avenues for future research in the more efficient inte- a specific application domain. One of the industries where
gration of AI in IoT. It raises possibilities for developing more efficient data flow and rapid response to anomalies are critical
advanced autonomous systems and continuous improvement is monitoring machinery and environmental conditions in
in managing large volumes of data. This lays the foundation industrial environments.

VOLUME 12, 2024 40369

 W. Villegas-Ch et al.: Toward Intelligent Monitoring in IoT: AI Applications

B. APPLICATION ENVIRONMENT: INDUSTRIAL the model for this new configuration. The adaptation and
MONITORING retraining mechanism ensure that as more data is collected
The system is designed to monitor, in real-time, both and feedback is introduced, the model becomes more accurate
machinery and environmental conditions within an industrial and efficient in detecting and responding to anomalies.
environment. With the proliferation of Industry 4.0, pro- The proposed system provides a comprehensive solution for
duction processes’ efficiency, safety, and optimization have monitoring machinery and environmental conditions in an
become a critical factor [11]. Therefore, it is essential to industrial environment, ensuring more efficient, safe, and
have a robust and reliable system that can provide accurate optimized production.
and timely information on the state of the machinery and
environmental conditions, thus allowing rapid intervention C. REVIEW OF SIMILAR WORKS
in the event of anomalies and continuous improvement of The proliferation of IoT has caused an explosion in the
processes [12]. generation and transmission of data. Given the scale and
For this, several sensors are used, such as the temperature diversity of this data, significant efforts have been made
sensor, which monitors the ambient temperature and the in the research field to address and optimize this flow of
machines. Their primary function is to ensure that machinery information. Starting from the very essence of the problem,
does not overheat, which could lead to premature failure or several works have discussed the data volume challenge
interruptions in production. The humidity sensor is vital in in IoT. The work [19] noted that although IoT devices
monitoring humidity levels in an industrial environment [13]. generate a torrent of potentially valuable data, transmitting
This monitoring is crucial to prevent corrosion of machines all that data for centralized analysis is not always practical
and to ensure the quality of certain products that may be or desirable. This concern is reflected in the findings of [20],
sensitive to high or low humidity levels. For its part, the who identified the costs associated with constantly sending
pressure sensor is used to monitor machines that operate large amounts of data to the cloud regarding bandwidth and
under certain pressure levels. These machines include but are finances.
not limited to, boilers or hydraulic systems. An unusual vari- Given the understanding of these challenges, multiple
ation in pressure may indicate faults or non-optimal operating proposals exist to address them. One of the most popular
conditions. The primary function of the Accelerometer is to approaches is edge processing, where data is processed
detect unusual vibrations in machinery. When seen in time, locally, close to where it is generated, before deciding
these vibrations can indicate wear, malfunction, or the need what to transmit. According to [21], this approach not
for preventive maintenance [14]. The IoT Camera is used for only reduces the data load on the network but can also
visual monitoring of the production line. This supervision provide real-time responses, which is essential for critical
allows for early detection of obstructions, failures, or any applications. However, edge processing has its limitations.
other problem that may arise, thus guaranteeing smooth As noted in [22], while edge processing can alleviate some
operation and minimizing downtime in production [15]. challenges, it is not a panacea. IoT devices, especially those
The factory deployment process involves strategically designed to be energy-efficient and cost-effective, may lack
placing these sensors on critical machines and specific areas the computational power needed to perform sophisticated
of the production environment. The placement of these analytics.
devices must be such that they can provide a panoramic This becomes the basis of our proposal, implement-
view of the operational and environmental health of the ing AI-based adaptive systems. Several researchers have
factory [16]. Once the sensors collect the data, it is transmitted explored the possibility of lightweight machine learning
to the central server, where the AI model processes and models at the intersection of AI and IoT. For example,
analyzes it in real time. If any anomaly is detected, alerts can [23] proposed a machine learning algorithm adapted for
be activated for factory personnel to take corrective action. resource-constrained devices. However, while this research
Optimizations made to ensure real-time efficiency are crucial has demonstrated the viability of AI in IoT, there is still a
in an industrial environment [17]. Since factories operate in gap in adaptive systems that can learn and evolve in real-time
continuous production cycles, any interruption or delay can based on the data they receive.
result in significant losses. Therefore, the system is designed The gap in the existing literature is the lack of systems that
to work with minimal latency and quickly respond to detected can filter and categorize data and adapt to changing patterns
situations. in real-time. Adaptability is crucial, as the environment in
In an industrial environment, conditions are constantly which IoT devices operate is dynamic, and patterns observed
changing. Machines wear out, production patterns vary, and one day may not be relevant the next. Our proposal seeks
external factors, such as weather conditions, can influence the to fill this gap. While we leverage accumulated knowledge
internal environment. Therefore, it is vital that the system on edge processing and lightweight AI models, we focus on
can adapt to these changes. The procedure for introducing adaptability [24].
simulated feedback allows the model to adjust to new con- The review of previous works has identified that, although
ditions [18]. If, for example, a new machine is installed or a numerous systems are dedicated to real-time monitoring
production line is modified, the feedback will help recalibrate and predictive analysis, few have effectively integrated
40370 VOLUME 12, 2024
W. Villegas-Ch et al.: Toward Intelligent Monitoring in IoT: AI Applications

adaptability and continuous learning capabilities, as proposed optimized to run on Raspberry Pi devices. We have used
in our system. In particular, existing systems [25], [26] Python, a versatile programming language with various
offer solutions based on IoT and AI technologies but focus libraries for machine learning and data processing, such as
primarily on data collection and presentation without fully TensorFlow, Keras, and Scikit-learn, for implementing our
realizing the potential of deep predictive analysis and real- intelligent system [31], [32].
time adaptability, unlike System [25], which is limited to
standard monitoring functionality without adaptive capa- E. DEPLOYMENT ENVIRONMENT
bilities, and System [26], which, although incorporates AI At the forefront of contemporary technological development,
elements, does not optimize energy efficiency or scalability. the ability to manage and process large volumes of data in
By developing a system that can learn from the data it real time is essential. This work focuses on designing and
receives, adapt to changes, and make informed decisions implementing a natural environment that efficiently manages
about what to transmit and when, we aim to offer a more the interactions and operations of IoT devices.
complete and robust solution to the data volume challenge The infrastructure is based on a high-performance central
in IoT [27]. Although the existing literature has established server that acts as the core of our system. This server has an
a solid foundation on which to build, our proposal for Intel Xeon E5-2620 v4 processor, 64 GB of RAM, and a 2 TB
an AI-based adaptive system offers a significant advance, SSD. The powerful hardware ensures fast data management
offering a more dynamic and evolutionary solution to the and is the main point of contact for the Raspberry Pi 4 Model
constant challenge of data management in IoT devices [28]. B instances that we have deployed in the field as active IoT
nodes.
D. IoT DEVICES The operating system chosen for this environment is
The IoT devices used in this study are essential to collect, Ubuntu Server 20.04 LTS. This choice is based on Ubuntu’s
process, and transmit data in real-time. For this research, reliability, stability, and wide range of tools for managing
we have selected the Raspberry Pi due to its processing and monitoring complex systems [33]. The implementation
power, connectivity, and flexibility to implement machine and functionality of the system depends mainly on the
learning systems [29]. tools and libraries we use:
• Python: Our primary programming language, used for
1) DESCRIPTION AND TECHNICAL SPECIFICATIONS
its versatility and rich library offering. Python takes care
The IoT device chosen for this study is the Raspberry of the primary operations, from communicating with
Pi 4 Model B. This device is widely recognized for its devices to analyzing data.
adaptability in various applications, from weather stations to • Pandas: An indispensable tool for managing and analyz-
surveillance systems [30]. The Raspberry Pi 4 Model B has ing data in our environment. It allows us to process the
multiple technical characteristics that make it ideal for this data we receive from IoT devices quickly and efficiently.
study: • TensorFlow: Used for machine learning and artificial
• CPU: Quad-core Cortex-A72 at 1.5GHz.
intelligence. It is the heart of our adaptive data filtering
• GPU: Broadcom Video Core VI.
system, which runs on each IoT node.
• RAM: 2 GB, 4 GB, or 8 GB LPDDR4 options.
• MQTT: Communication between our IoT devices and
• Storage: microSD card for operating system and data
the server is carried out through the MQTT protocol,
storage.
given its efficiency and adaptability for IoT systems.
• Connectivity: WiFi 802.11ac, Bluetooth 5.0, Gigabit
• Docker: To guarantee the consistency and portability of
Ethernet.
our implementations, we use Docker, which allows us to
• Ports: 2 USB 3.0 ports, 2 USB 2.0 ports, GPIO, two
encapsulate our applications and services in containers.
micro-HDMI ports.
Connected Sensors: The proposed architecture for the deployment environment
• Temperature sensor: This module has a range of −40◦ C has been designed to optimize efficiency in data collection,
to 85◦ C and a precision of ±0.5◦ C. transmission, and analysis in an IoT ecosystem. This
• Humidity Sensor: Operating with a range of 0% to architecture, as illustrated in Figure 1, is a simplified but
100%, this sensor has an accuracy of ±2%. robust representation of how different devices in the IoT
• Pressure sensor: With an operating range of 300hPa to environment interact with a central server. Each IoT device
1100hPa, this sensor offers an accuracy of ± 1hPa. connects to the central server using WiFi signals or an
• 3-axis accelerometer: This module detects motion and Ethernet cable, depending on the connection type [34]. It is
can identify orientation and vibrations. vital to show that each device has its independent connection
to the server, highlighting the flexibility and scalability of the
2) MODEL AND VERSION OF SOFTWARE/FIRMWARE network.
The Raspberry Pi 4 Model B software for this study is the Directional arrows next to the connections will indicate
Raspberry Pi OS (previously known as Raspbian), version the data flow. These arrows can be two-way, showing that
‘‘Buster.’’ This Debian-based operating system has been devices send data to the central server and receive instructions

VOLUME 12, 2024 40371

 W. Villegas-Ch et al.: Toward Intelligent Monitoring in IoT: AI Applications

FIGURE 1. General architecture of the implementation environment.

or software updates from the server. The representation of this controlled environments and the other in industrial produc-
bidirectional flow is crucial to understanding how the system tion environments. In the controlled environments, data was
can adapt and respond in real-time to different situations. generated using a set of specific algorithms that mimic the
• Central Server: This is the backbone of our entire operating conditions and behavioral patterns of IoT devices.
architecture. The central server receives, processes, and This data was collected with a predefined sampling rate and
stores data from all connected IoT devices. Additionally, adjusted to simulate operating scenarios and system stress
the AI model we have developed to filter and categorize conditions.
data in real-time resides on this server. Every time one of IoT sensors are used in industrial environments. These
the sensors detects an anomaly or a significant change in sensors provided data on critical operational parameters,
the data it is monitoring, that information is transmitted including thermometers, barometers, and accelerometers.
to the central server for further analysis. This server must Data was collected continuously over 12 months, ensuring
have high processing and storage capacity, as it handles a representative sample that reflects seasonal variations and
a large amount of data in real-time. changing operating conditions.
• IoT Devices: Temperature Sensor: Monitors tempera- The size of the data sets and the temporal range selected
ture variations in the environment. Given its accuracy for this study were determined based on several key factors.
of ±0.5◦ C, it is susceptible to changes and can identify First, the wide temporal range of 12 months for sensor data
fluctuations that might go unnoticed by less precise and three months for controlled data was chosen to ensure
sensors. a comprehensive understanding of patterns and behaviors
• Humidity Sensor: This device is essential in environ- under different conditions and periods. This is crucial to
ments where the humidity level can affect the operation validate the robustness and adaptability of the proposed
of other devices or the quality of the product, for system. The volume of data collected was adjusted to capture
example, in agricultural or industrial storage. sufficient variability and complexity, ensuring that the AI
• Pressure Sensor: It is mainly used to detect variations system was trained and evaluated effectively. This data allows
in atmospheric pressure, which can be crucial in for a comprehensive system evaluation and assesses its
meteorological or altitude applications. accuracy and efficiency in standard and extreme conditions.
• Accelerometer: This three-dimensional sensor is essen-
tial to detect movements or vibrations. In industrial 1) DATA GENERATED BY SENSORS
applications, it can be used to monitor the health of The data comes directly from the IoT sensors integrated into
machines and detect anomalies in their operation. the system. These sensors, when active, collect information
• IoT Camera: Unlike traditional sensors, this smart from their environment and transmit this data to the central
camera can transmit images or videos to the central server at regular intervals.
server. Its application could be in security systems or Data set characteristics:
real-time process monitoring. • Duration: Data has been collected over 12 months.
Each device is connected to the central server, whose • Number of samples: Approximately 1 million records.
primary function is to collect data from the environment. Value range:
However, our proposal lies in the ability of these devices to • Temperature sensor: −40◦ C to 85◦ C
perform an initial ‘‘filtration’’ of data, thus reducing the load • Humidity sensor: 0% to 100%
on the network and optimizing transmission. • Pressure sensor: 300hPa to 1100hPa
• Accelerometer: Vector data in 3 axes.
F. DATA SET • IoT Camera: 1080p images.
For this solution’s implementation and validation process, During the collection period, several anomalies were
two approaches were used for data collection: one in observed in the data due to environmental factors, sensor

40372 VOLUME 12, 2024

W. Villegas-Ch et al.: Toward Intelligent Monitoring in IoT: AI Applications

failures, or external interference. These anomalies appear • One-Hot Encoding: Categorical features were trans-
in approximately 0.5% of total records and can vary in formed using one-hot encoding, representing them as
magnitude and frequency. binary vectors.
Table 1 offers a concise overview of the main preprocess-
2) GENERATED DATA FOR TESTING ing techniques applied to the data set. The description of
This data was created using specialized tools and libraries each technique and the specific tools or methods used for its
that simulate precise data characteristics and behaviors. implementation are highlighted. This information is a quick
Its primary goal is to provide a controlled means to test reference to understand the transformations performed and
and validate algorithms and processing techniques without ensure that the data is consistent, clean, and ready for deeper
affecting or relying on physical sensor data. analysis.
Data set characteristics:
• Duration: Simulation over three months. TABLE 1. Summary of pre-processing techniques used.
• Number of samples: Approximately 25,000 records.
• Range of values: Like the actual data set, but with the
possibility of entering extreme or out-of-range values to
evaluate the system’s robustness.
Since this dataset is for testing, anomalies were intention-
ally introduced at a higher frequency, around 5% of the total
records. These anomalies were designed to represent possible
failure or interference scenarios the system could encounter in
a real-world environment. Combining these two data sets has
allowed for more robust development and validation of our
solution, ensuring it can handle both normal and abnormal
conditions that may arise in a real-world environment. The
distinction between natural and artificial data confirms that 2) METHODOLOGY AND DESIGN OF THE AI MODEL FOR
the system is tested in various scenarios and conditions. THE FILTERING SYSTEM
G. METHOD Incorporating AI in IoT systems has revolutionized how data
1) DATA PRE-PROCESSING is managed and processed. We have designed an AI-based
filtering system to address the challenges presented by
Data preprocessing ensures that data is ready for analysis
and modeling. Various techniques and tools are applied to real-time data management and transmission [36]. Proper
clean, transform, and prepare the data sets. The objective is design of an AI model is essential to ensure accuracy,
to improve the data quality and facilitate its subsequent use efficiency, and robustness in addressing specific problems.
in the system. The choice of the model, the parameters, and the training
The techniques and tools used to clean and prepare the data procedure are decisive in the system’s effectiveness [37],
set are: [38].
• Detection and elimination of outliers: We use statistical The core of the system focuses on machine learning
methods such as the interquartile range (IQR) rule to models. We chose a combination of recurrent neural networks
identify and eliminate values that deviate significantly (RNN) and decision trees. RNNs specialize in processing
from the general behavior of the data. sequential data, such as time series, making them ideal for
• Handling of missing values: Imputation was used using detecting patterns in data that occur over time [39]. For
the mean, median, or mode, depending on the nature and their part, decision trees are responsible for filtering specific
distribution of the data in the specific column. events based on predefined criteria. We train these models
• Tools: For these tasks, specialized Python libraries are using vast historical data sets from IoT devices to ensure
used, such as Pandas and Scikit-learn, which offer com- system adaptability and effectiveness in different scenarios.
prehensive functionalities for data pre-processing [35]. RNNs have the particularity of processing data sequences,
Transformations or normalizations applied. maintaining a ‘‘memory’’ of previous data in the sequence,
• Standardization: Numerical data was standardized to which is essential for understanding contexts and detecting
have a mean of 0 and a standard deviation of 1. patterns in time series [40].
This facilitates comparison between different features Among the model parameters and architecture, the RNN
and improves the convergence of machine learning architecture consists of several hidden layers that process and
algorithms. transform the input data before producing an output.
• Min-Max Normalization: In some cases, Min-Max • Input Layer: Receives data from each sensor and
normalization was used to scale numerical features to prepares it for processing in subsequent layers. Its size
the range [0, 1]. This technique is beneficial when the depends on the number of input features.
data has very different ranges of values, and you want to • Hidden Layers: The middle layers do most of the
maintain the relation-ship between the values. processing. Our model uses three hidden layers, each

VOLUME 12, 2024 40373

 W. Villegas-Ch et al.: Toward Intelligent Monitoring in IoT: AI Applications

with 128, 64, and 32 neurons, respectively. Each The integration process was carried out in several stages;
layer uses the ReLU (Rectified Linear Unit) activation once trained, the DNN model was converted to a format
function due to its efficiency and ability to reduce the more suitable for devices with limited resources. We use
gradient fading problem. tools like TensorFlow Lite, which allows you to convert
• Output Layer: Produces the final prediction of the TensorFlow models into versions optimized for edge devices.
model. Since our goal is anomaly detection, the output The converted model was integrated into the firmware of the
layer has only one neuron with a sigmoid activation IoT device. This involved writing specific code that loaded
function, which gives a probability that a particular input the model into the device’s memory, processed the input
is anomalous. data, and ran inferences using the model. Next, an interface
Training an RNN is an iterative process that adjusts the was established so the model could receive data from the
weights and biases of the network to minimize the error sensors and send results to activate some device functions or
between the model predictions and the actual values. For communicate with a central server.
this purpose, we use the Adam optimizer. This optimizer Given the nature of IoT devices, efficiency is paramount.
combines the salient properties of other popular optimization For this, several optimizations were implemented to ensure
algorithms, ensuring fast convergence and stable perfor- real-time performance:
mance. We set an initial learning rate of 0.001. Although • Pruning: This technique eliminates weights or neurons
Adam can adjust this rate automatically throughout training, from the model that have a minimal impact on accuracy.
establishing an appropriate starting point is essential. The This significantly reduces model size, resulting in faster
model was trained for 50 epochs. An epoch refers to a inference times.
complete traversal of the entire data set. Given the volume • Quantization: We convert the weights and biases of the
of our dataset, we determined that 50 epochs were optimal to 32-bit model to lower-bit formats, such as 8 bits. This
ensure model convergence. reduces the model size and speeds up inference due to
We employ binary cross entropy as our loss function lower computational load.
when addressing a binary classification problem, such as • Runtime Model Compilation: We use specialized com-
anomaly detection. This function evaluates the discrepancy pilers that translate the model into device-specific
between model predictions and actual values, making it the instructions, optimizing performance.
ideal choice for binary classification problems. To mitigate • Memory Optimization: We ensured that the model
the risk of overfitting, especially in the case of RNNs, loaded into the fastest memory available on the device,
we implement regularization techniques, such as dropout. such as SRAM, instead of flash or external memory.
This technique randomly deactivates specific neurons during • Batch Processing: Wherever possible, we processed
training, forcing the network to generate more general data in batches rather than individually, which takes
and robust representations [41], [42]. Opting for a Neural advantage of parallelization and improves efficiency.
Network, whether Deep or Recurrent, with its respective With these optimizations, the model works in real-time on
parameters and training techniques reinforces our objective of the IoT device, ensuring that inferences are made promptly
detecting anomalies in an IoT context, guaranteeing precision and accurately. These adaptations are essential to ensure that
and efficiency in our predictions. implementing AI in IoT devices is feasible and effective in
The system was designed with two essential criteria: real-world scenarios [43].
• Redundancy: Repeated data or data without significant
changes in each period are considered redundant and 4) VALIDATION AND TESTING
filtered out. Validation and testing are essential to ensure the proper
• Priority: Levels are established for different types of functioning of any system, especially in applications where
data. Critical alerts or significant changes have high AI plays a crucial role. In our proposal, the successful
priority, while routine readings are grouped for less implementation of the model on the IoT device requires that
frequent transmissions. the model produces accurate results and operates efficiently
After evaluating the data based on the criteria, the system in real-time and various scenarios. For the controlled envi-
decides in milliseconds whether to transmit, group, or discard ronment testing scenario, the IoT device with the integrated
a piece of data. This real-time capability ensures the proper model is tested in a laboratory or similar environment, where
response of IoT devices to changing conditions. all variables and conditions can be controlled and monitored.
This lets you evaluate the system’s performance under ideal
3) DEVICE DEPLOYMENT conditions and detect fundamental problems or apparent
Implementing the AI model on IoT devices represents a sig- failures. For example, it is possible to simulate a series of
nificant challenge, mainly due to the resource limitations of specific temperature and humidity values and check whether
these devices compared to traditional servers or workstations; the device provides the expected responses.
the process is presented in Figure 2. The main objective is to In the second real-simulated environment, the device is
ensure that the model operates efficiently in real time without tested in an environment that simulates the actual conditions
compromising the accuracy or functionality of the device. in which it is expected to operate but still can monitor and

40374 VOLUME 12, 2024

W. Villegas-Ch et al.: Toward Intelligent Monitoring in IoT: AI Applications

FIGURE 2. Model integration process in the IoT devicemodel integration process in the IoT device.

control some variables. This provides a more realistic view of • F1-Score: A metric combines precision and recall
system performance and helps identify problems that might into a single value. It is beneficial when classes are
not be evident in a controlled environment. One scenario unbalanced.
could involve deploying the device in a factory or smart home
for a short period. Precision × Recall
To evaluate the performance of the system, several metrics F1 = 2 × (4)
Precision + Recall
were used that provide valuable information about different
aspects of the system: The choice of specific metrics depends on the exact nature
• Accuracy: This fundamental metric measures the pro- of the problem and the consequences of different types of
portion of correct predictions made by the model about errors. For example, recall might be more important than
the total predictions. precision in an application where detecting all possible
Number of correct predictions anomalies (even if some false positives occur) is crucial.
Accuracy = (1) The validation and testing process is essential to ensure
Total number of predictions
our system performs as expected under various conditions.
• Latency: Since it is being implemented in IoT devices Using a combination of testing in controlled and simulated
where accurate time is of the essence, latency (the environments and a series of mathematically sound metrics,
time it takes for the system to process an input and we can be confident that our system is robust, efficient, and
produce an output) is crucial. This metric is measured accurate.
in milliseconds (ms); the lower, the better. Our focus on validation and testing ensures proper func-
• Resource usage: It is vital to know how much resource tionality and system effectiveness, especially in applications
(such as CPU, memory, and power) the system uses, where AI plays a central role. A set of metrics, including
especially on devices with limited resources [44]. This accuracy, latency, and resource usage, have been applied
metric can be a percentage or an absolute value, to evaluate the performance of our model in realistically
depending on the nature of the resource. simulated and controlled environments. However, measuring
• Recall: This metric measures the proportion of true pos- energy efficiency, scalability, robustness to interruptions, and
itives that were correctly identified. It is essential when other aspects is essential to deepen the understanding of
the consequences of a false negative are significant. comparative performance. By integrating these parameters,
True positives a more holistic and detailed view of the contribution of the
Recall = (2) proposed system will be provided.
True positives + False negatives
Parameters include:
• Precision: Precision measures the proportion of correct
identifications. • Energy Efficiency: Measurement of system energy
consumption during normal and high-load operations.
True positives This parameter is crucial for IoT systems that operate
Precision = (3)
True positives + False positives with limited resources.

VOLUME 12, 2024 40375

 W. Villegas-Ch et al.: Toward Intelligent Monitoring in IoT: AI Applications

• Scalability Capacity: Evaluation of the system’s capac- A. SYSTEM PERFORMANCE IN A CONTROLLED

ity to handle increases in the number of IoT devices and ENVIRONMENT
the volume of data processed. A set of data generated in a controlled environment within
• Disruption Robustness: Determination of the system’s the genuine factory was used for the initial validation of the
ability to maintain stable operations in the face of system. These data provided a solid basis for measuring the
network interruptions or hardware failures. accuracy and efficiency of the system under ideal conditions.
• Downtime: Measurement of the time required to recover Table 2 shows the measurements captured by the tem-
the entire operation after failures or interruptions. perature sensor over a 10-hour interval. These readings
• Interoperability: Assessment of the system’s ability to demonstrate the system’s ability to track and document minor
integrate and operate with various IoT devices and variations in the temperature of the controlled environment.
systems. The recorded values indicate that temperatures vary between
• Total Operation Cost: Analysis of the cost of imple- 22.1◦ C and 24.9◦ C. It is crucial to note that a previously
mentation and maintenance of the system, including established limit was exceeded at four specific times (hours
hardware, software, and operations expenses. 2, 5, 7, and 10), triggering corresponding alerts. These
exceedance episodes highlight the accuracy and efficiency of
5) FEEDBACK AND CONTINUOUS LEARNING the system in identifying critical temperature changes, reach-
ing an accuracy rate of 98.7%. This precision rate reaffirms
In the modern era of AI, a system’s ability to adapt and
the system’s reliability in the temperature monitoring task.
learn from new information is a distinctive characteristic
It suggests its strength in the face of possible variations,
that allows it to maintain its relevance and accuracy over
consolidating itself as a fundamental tool to prevent risks
time. In the context of IoT devices, where conditions and
associated with machinery overheating in industrial contexts.
environment can change, it is essential that systems not only
operate based on the data they were initially trained on but TABLE 2. Temperature sensor results.
also evolve based on feedback and new experiences.
Feedback is a technique used to test how the system
would behave when faced with new data or scenarios without
waiting for them to occur in real situations. To do this,
simulated data is generated using specialized tools and
algorithms; data sets that imitate the characteristics and
behaviors of accurate data are created. These sets can include
hypothetical or extreme scenarios the system has not yet
encountered.
For incorporation into the system, this data is entered into
the system as if it were actual data. The system processes
Table 3 details the measurements acquired by the humidity
this information and makes decisions based on it. The
sensor over 10 hours. The measurements reflect the system’s
system response is compared with the expected responses.
ability to discern and document small changes in the humidity
Discrepancies are recorded and analyzed to understand how
of the controlled environment. Humidity levels fluctuate
and why the system made certain decisions. Based on the
between 44% and 70%. Notably, at three specific times
comparison, feedback points are generated. This feedback
(hours 5, 6, and 8), a previously defined limit was exceeded,
tells the system whether its answer was correct and what it
triggering alerts. These exceedance episodes highlight the
should have done differently if it was incorrect.
accuracy and effectiveness of the system in recognizing
critical deviations in humidity, achieving an accuracy rate
III. RESULTS of 97.8%. This degree of accuracy suggests that the system
With the technological advancement and increasing inte- is highly reliable for humidity monitoring. It denotes its
gration of AI systems in the industry, the present study robustness in the face of changes, becoming an essential
develops an AI-based adaptive system for IoT devices that resource to mitigate possible problems associated with
monitor machinery and environmental conditions in a factory. humidity in an industrial plant, such as corrosion or damage
The main objective was to design a system that could to the quality of certain products.
efficiently discern between critical data that had to be Table 4 shows the pressure sensor readings collected over
transmitted immediately and those that could be retained or 10 hours. These measurements demonstrate the system’s
processed locally, all to optimize bandwidth and guarantee ability to detect and record variations in the pressure
rapid responses to abnormal events. This approach is based of the controlled environment. The pressure values range
on the central hypothesis that integrating AI into IoT devices between 100875 Pa and 101325 Pa. It is crucial to mention
directly at the point of data collection can reduce network that a predetermined threshold was crossed in three specific
overhead and improve the speed and accuracy of identifying instances (hours 6, 7, and 9), activating alerts. These alerts
critical events or patterns. highlight the system’s precision by identifying significant

40376 VOLUME 12, 2024

W. Villegas-Ch et al.: Toward Intelligent Monitoring in IoT: AI Applications

TABLE 3. Humidity sensor results. TABLE 5. Accelerometer sensor results.

alterations in pressure, with a success rate of 96.9%. This the device’s ability to monitor the production line efficiently,
level of accuracy validates the reliability of the pressure although they also point out potential areas for improvement
monitoring system. It suggests that it is robust against in anomaly detection. Any industrial monitoring system must
alterations, a crucial instrument to guarantee the safety and maintain a high detection rate to ensure the continuity and
efficiency of equipment operating under certain pressure quality of production and minimize associated risks.
levels, such as boilers or hydraulic systems.
TABLE 6. Humidity sensor results.
TABLE 4. Pressure sensor results.

Figure 3 shows a composite set of graphs representing key

Table 5 shows the measurements taken from the metrics monitored over 10 hours in a controlled environment.
accelerometer over 10 hours. This data reflects the vibrations In the first subgraph, the temperature remained in a narrow
detected in the controlled environment, expressed in meters range but with three notable peaks at hours 2, 5, and 10,
per second squared (m/s2 ). Vibrations in industrial machinery where the preset threshold was exceeded. The humidity
are a crucial indicator of its condition and operation. subgraph shows a relatively stable pattern with some minor
An unusual increase in these may signal premature wear fluctuations. The pressure in the third subgraph also showed
or a problem in the operation of the machine. During the fluctuations, particularly at hours 3, 6, and 9. The vibrations
recorded period, it is observed that the vibrations ranged of the accelerometer in the fourth subgraph indicate three
between 0.002 m/s2 and 0.008 m/s2 . At hours 2, 5, and 9, there points of interest at hours 2, 5, and 9, where the readings
were three specific moments where the vibrations exceeded a exceeded a certain threshold. Finally, the fifth subgraph
predefined threshold, causing the system to generate alerts. visually represents the number of failures detected by the IoT
These measurements and alerts underscore the system’s camera in each hour, with peaks at hours 2, 5, and 8.
ability to detect subtle but significant variations in vibrations, The analysis was expanded to include a set of additional
achieving high precision. This tool is, therefore, presented metrics to provide a comprehensive and quantifiable assess-
as essential in the early detection of possible failures or ment of system performance. These metrics allow us to
breakdowns in industrial equipment. address the critical dimensions of performance, efficiency,
Table 6 details the observations for the IoT Camera reliability, and user satisfaction, allowing us to compare
over 10 hours. This table compares the faults intentionally with existing systems. Table 7 presents a summary of these
introduced into the system and the faults that the camera results. This comparative analysis highlights the robustness
detected. A high coincidence can be observed between the and efficiency of our system in intelligent monitoring and
faults introduced and detected throughout the recorded hours. predictive analysis.
However, at specific times, such as the second and eighth, The results obtained in the table reveal several critical
a slight discrepancy is noted between the errors introduced dimensions in which our proposed system outperforms
and those identified by the camera. In total, 23 bugs were existing solutions, marking a significant advance in real-time
submitted, and the camera detected 21 accurately. This monitoring and analysis technology. The improved through-
translates into a success rate of 91.3%. These data reaffirm put indicates a superior capacity to handle large volumes

VOLUME 12, 2024 40377

 W. Villegas-Ch et al.: Toward Intelligent Monitoring in IoT: AI Applications

FIGURE 3. Integrated monitoring of parameters in controlled environment.

TABLE 7. Comparison of the performance of the proposed system against environment presents external factors, interferences, and
existing systems in a controlled environment.
other elements that could affect the system’s performance.
Still, it keeps certain variables under control to analyze the
behavior systematically.
It is essential to compare the results visually to under-
stand the differences between performance in a controlled
environment and the real simulated one. Figure 4 pro-
vides a clear and comparative view of how the system
monitors and records temperature and humidity in two
different scenarios: a controlled environment and a simu-
lated natural environment. At first glance, we can identify
of data, essential in IoT environments where speed and specific trends and critical characteristics. For example, the
efficiency in information processing are critical. The cost curve corresponding to the temperature in the controlled
efficiency highlights the economic viability of our system, environment, represented by the blue line, shows a rel-
demonstrating that it is possible to achieve high levels of atively stable variation. This suggests that the conditions
performance without incurring prohibitive costs. The low in that environment were consistent and that the system
error rate emphasizes the reliability and precision of the sys- could operate without interruptions or disturbing external
tem, fundamental aspects for adopting artificial intelligence factors.
technologies in critical applications. Furthermore, high user The orange line, which represents the temperature in
satisfaction validates the system’s usability and effectiveness the simulated natural environment, shows more significant
from the end-user perspective and indicates an alignment with fluctuations compared to the controlled environment. This is
market needs and expectations. indicative that, although the environment attempts to replicate
natural conditions, there are still controlled variables that
B. SYSTEM PERFORMANCE IN A CONTROLLED affect sensor performance. More pronounced oscillations
ENVIRONMENT could result from variations in heat sources or changes in
For the proposed monitoring system based on IoT sensors, airflow. Looking at the green curve, we notice stability in
after validating its performance in a strictly controlled envi- the humidity readings of the controlled environment. This
ronment, a series of tests were carried out in an environment indicates that the environment remained relatively constant
that simulated actual conditions. This simulated natural in humidity, providing a solid basis for testing.
40378 VOLUME 12, 2024
W. Villegas-Ch et al.: Toward Intelligent Monitoring in IoT: AI Applications

In contrast, the red line, which represents the humidity alerts issued. However, when analyzing the results of the
in the simulated natural environment, shows more notable simulated natural environment, an occasional divergence is
variations. These fluctuations could be due to factors such perceived between the recorded vibration magnitudes and
as the presence of steam-generating machines or changes the activated alerts. These discrepancies suggest that the
in environmental conditions. It is essential to point out the accelerometer, faced with more variable and less predictable
points where the curves, both for temperature and humidity, conditions, may have experienced disturbances or external
converge or diverge significantly. These points represent influences that affected its ability to discern when it was
moments where the system reacted similarly or differently appropriate to trigger an alert.
in both environments, which is crucial to determine its It is essential to mention that while the deviations observed
robustness and reliability. in the simulated natural environment may be of concern, they
Figure 5 compares the pressure sensor readings in two also offer an invaluable opportunity to calibrate and adjust
environments: the controlled one and the simulated real the accelerometer’s alert threshold. Recognizing these areas
one. There are some discrepancies in how the sensor of potential improvement contributes to the adaptation and
operates under these two circumstances. In the controlled refinement of the system, preparing it to face more complex
environment, readings tend to be more stable, suggesting an challenges with greater precision and reliability.
environment without significant disturbances. On the other
hand, the simulated natural environment shows more pro-
C. INTEGRATED ANALYSIS OF PARAMETERS AND
nounced fluctuations, which could indicate natural variations
TRANSMISSION EFFICIENCY
or interventions in the system that simulate real situations.
Table 8 presents the parameters monitored over 10 hours.
The contrast between these two curves is essential to
When looking at the measurements, several key points can be
evaluate the robustness of the sensor. While variations are
highlighted. During the observed period, a gradual increase
expected in a more dynamic environment, it is crucial to
in temperature was recorded, reaching a peak of 25.2◦ C in
identify any significant deviations that could compromise
the 5th hour. This increase could be due to external factors
sensor accuracy. The proximity or divergence between the
not controlled during the study or an increase in system
two lines at specific times can point to areas for improvement
activity. However, the temperature decreased from the 6th
or recalibration.
hour onwards, suggesting a stabilization process or a response
Figure 6 presents the performance of the IoT Camera
to the cooling system.
in fault detection, comparing the results obtained both
Humidity decreased over time, going from 50% in the
in the controlled environment and the simulated natural
first hour to 45% in the 10th hour. This indicates that
environment. By studying the trends, it is apparent that the
the surrounding conditions were progressively drier, which
camera demonstrates a consistent ability to identify failures
may influence other parameters or the overall system
in a controlled environment. However, there are notable
efficiency. Throughout the 10 hours, the pressure experienced
points where this consistency decreases in the simulated real
a constant, although subtle, decrease, going from 101325 Pa
environment, indicating possible challenges or interferences
to 101240 Pa. It is essential to determine whether this
that may arise in a more complex context.
behavior is related to atmospheric changes or the system’s
Furthermore, the detection pattern in the simulated real
internal dynamics.
environment shows certain irregularities, reflecting the pres-
Vibration measurements showed minimal variations, with
ence of unforeseen elements or variables that may have
a noticeable increase in the 9th hour. This spike may indicate
influenced the camera’s performance. These irregularities are
an abnormal event or disturbance in the system that should
essential to understanding how the camera could react to
be investigated further. Although most hours presented few
unplanned situations and how its accuracy could be improved
failures, an increase is observed in the 5th and 9th hours.
in dynamic contexts. Altogether, the figure reveals the
These spikes in failures could be correlated with notable
adaptability and potential limitations of the IoT Camera when
changes in temperature and vibration, respectively. Efficiency
faced with changing environments. While performance in the
remained high throughout the period, with an average value
controlled environment is a positive sign of its effectiveness,
close to 98%. However, it is vital to investigate the hours
the variations observed in the simulated real environment
where there was a slight decrease, especially in the 9th hour,
provide valuable insight into areas of improvement, optimiza-
where the efficiency dropped to 96%, to ensure the integrity
tion, and recalibration to ensure accurate and reliable fault
and reliability of the transmitted data.
detection in broader applications.
Figure 7 puts into perspective the performance of the
accelerometer by recording vibrations and generating rel- D. ANALYSIS OF PERFORMANCE METRICS
evant alerts, both in a controlled environment and in a To fully understand the behavior and efficiency of the imple-
simulated real environment. When inspecting the layout, it is mented system, it is essential to analyze the selected metrics
highlighted that the accelerometer maintains high sensitivity to evaluate its performance meticulously. Metrics offer a
and precision in the controlled environment, evidencing a quantitative view of the system and provide information about
direct correlation between the vibrations detected and the its qualitative behavior under different conditions.

VOLUME 12, 2024 40379

 W. Villegas-Ch et al.: Toward Intelligent Monitoring in IoT: AI Applications

FIGURE 4. Comparison of temperature and humidity in controlled and simulated environments.

TABLE 8. Confusion matrix built with the results obtained for the test standards and previous studies. The selected metrics cover
with 54 images.
aspects of both measurement accuracy and response time
and the incidence of errors. Notably, in most metrics, the
system not only meets the reference values but often exceeds
them. This is indicative of the robustness of the design
and implementation. Precisely, efficiency is measured in
percentages for a more direct interpretation. For example,
the accuracy of the Temperature Sensor reached 98.7%,
exceeding the benchmark value by 1.7%. These numeri-
cal details, presented in an organized manner, provide a
concise.
According to these results, it is evident that the accuracy
TABLE 9. System performance metrics. of all sensors slightly exceeds the benchmark values. This
is indicative of the robustness and reliability of the system.
Notably, the accelerometer and IoT camera, which could
be considered the devices most sensitive to environmental
variables, still exceeded the expected standards. The mean
response time, another crucial aspect in real-time moni-
toring systems, is also in line with accepted standards,
ensuring that any anomaly detected can be addressed
promptly.
As for false positives and negatives, they are within
acceptable margins. However, continuing to work in these
areas is vital to reduce these values further. A false positive
or negative in critical industrial environments can have
significant repercussions. While the system has exceeded
expectations on several metrics, it is important to recognize
Table 9 compares the performance of the system under areas for improvement. Although false positive and negative
study and the benchmark or reference values from industrial rates are within acceptable limits, reducing these numbers
40380 VOLUME 12, 2024
W. Villegas-Ch et al.: Toward Intelligent Monitoring in IoT: AI Applications

FIGURE 5. Comparison between pressure sensor results.

should be a priority, as they can influence decision-making fifth week, where accuracy rose to almost 99% by the twelfth
and operator confidence in the system. week. This notable increase highlights the importance of an
effective feedback system, demonstrating how the AI model
can quickly learn and adapt from actual data and contextual
E. RESULTS OF FEEDBACK AND CONTINUOUS LEARNING feedback.
Feedback is a vital tool in the evolution of any AI-based sys- The integration of this continuous feedback and learning
tem. In this case, it has been used with two main objectives: to system validates the system’s initial functionality and under-
validate the accuracy of the system’s predictions and to allow lines its evolution in terms of adaptation and accuracy over
its adaptation to new patterns and conditions. time. This ensures that the system not only operates based on
Feedback was systematically collected from responses the data it was initially trained on but also continually evolves
from users and technicians, who provided critical comments and improves its performance in the face of new experiences
on the alerts’ accuracy and the readings’ effectiveness under and changing conditions.
various environmental conditions. The generation of data The implemented AI-based filtering system was designed
sets complemented this process to represent hypothetical or to focus on continuous learning and adaptability. A clear
extreme scenarios the system had not yet encountered. pattern of improvement and optimization is observed by
This data was fed into the system as event data from focusing on how the system learned and adapted to incoming
the sensors, allowing it to process and make decisions in a data and user feedback. This phase not only validates the
controlled environment. The system’s response to this data initial functionality of the system but also highlights its
was then compared to the expected responses. Discrepancies evolution in terms of adaptation and accuracy over time.
identified in this process were analyzed to understand how A comprehensive feedback mechanism was incorporated,
and why the system made certain decisions, and based where responses and comments from users and technicians
on this, specific feedback points were generated. These on alerts, accuracy of readings, and efficiency under vari-
feedback points informed the system whether its answer ous environmental conditions were meticulously collected.
was correct and, if incorrect, what it should have done Subsequently, this information was reintroduced into the
differently. system, allowing the model to be refined and calibrated for
The impact of this approach on system learning and adapt- a more precise adaptation to real-time conditions. Figure 8
ability was evident through significant performance improve- graphically illustrates the progression and improvement of
ments. With the introduction of feedback, we observed a model performance with the systematic incorporation of
steeper increase in model accuracy, especially starting in the feedback.
VOLUME 12, 2024 40381
 W. Villegas-Ch et al.: Toward Intelligent Monitoring in IoT: AI Applications

FIGURE 6. Comparison of fault detection with the IoT camera.

The Figure shows the evolution of the model’s perfor- 1) NOTABLE SUCCESS STORIES
mance with and without feedback over 12 weeks. The In one situation, the system identified a gradual increase
model’s initial accuracy, represented by circles, shows a in the temperature of specific machinery before it reached
gradual increase, going from 92% in the first week to dangerous levels. This early detection allowed immediate
95.5% in the twelfth week. However, it is notable that, intervention, avoiding potential failures and guaranteeing
by introducing feedback, represented by the marks in operability. For example, for one night, the IoT camera
the shape of ‘x,’ the model’s performance experiences started recording inconsistent data due to a lighting defect.
a more pronounced increase. Starting at week 5, the Upon detecting these anomalies, the tool sent an alert in
model benefits significantly from the feedback provided, real time, facilitating the rapid resolution of the problem
and its accuracy skyrockets, reaching almost 99% by and guaranteeing visibility in the area in question. For
week 12. their part, accelerometer measurements exposed specific
The yellow shaded area starting at week 5 highlights patterns in certain machines during their startup phase. Once
the period in which significant feedback was introduced, these patterns were recognized, operational adjustments were
visually showing its positive impact on improving model made, resulting in optimal performance and equipment life
performance. This sharp increase underlines the impor- extension.
tance of an effective feedback system, demonstrating how
the AI model can quickly learn and adapt from sen-
sor data and contextual feedback. Ultimately, this anal- 2) RELEVANT FINDINGS AND LEARNING
ysis reinforces the idea that a well-structured feedback Through continuous monitoring, the system identified regular
mechanism is crucial to optimize and adjust AI-based fluctuations in humidity levels depending on the time of
models. day. This information has been vital for the optimization
of specific processes, especially those that are sensitive to
F. SUCCESS STORIES AND RELEVANT FINDINGS changes in humidity. Additionally, it was observed that the
Implementing an AI-based filtering system has proven to be sensors had slight delays in their readings when exposed
an effective tool not only for real-time monitoring but also to elevated temperatures. Based on this discovery, specific
for proactive identification of patterns and anomalies that calibrations were made, increasing the system’s overall
can directly impact operation and efficiency. Below are some accuracy. Likewise, the system could detect specific vibra-
notable success stories and insightful findings obtained using tional patterns that coincided with periods of maintenance
this tool. work or external alterations. This information has proven
40382 VOLUME 12, 2024
W. Villegas-Ch et al.: Toward Intelligent Monitoring in IoT: AI Applications

FIGURE 7. Comparison of alerts generated by the accelerometer.

crucial to properly schedule interventions and maintenance Another telling indicator is the system’s ability to detect
properly, thus minimizing interruptions and guaranteeing anomalies. In the case of the machine that showed an
greater efficiency. increasing trend in temperature, our system provided a crucial
alert when the temperature variation was only 1.5%. This
IV. DISCUSSION contrasts with traditional systems, which typically require
The practical applications of our work are vast and far- up to 5% variations before generating an alert. Furthermore,
reaching. From improving manufacturing efficiency to from an academic point of view, the pace of adaptation
optimizing critical infrastructure management, our approach and learning of the system is notable. The pronounced
provides a vital decision-making tool based on accurate, real- improvement in accuracy after the fifth week (from 95.2%
time data. For future work, it is considered that this research to 96%) reflects feedback’s incorporation and effectiveness.
not only expands current knowledge but also suggests new Comparing this adaptation rate with similar studies reveals
directions to explore, both in improving existing technology that our system adapts almost 50% faster to new conditions
and creating new interaction paradigms between AI and IoT. and feedback than most existing proposals [53].
Quantifying performance is essential when optimizing Our system demonstrates significant improvements in
real-time operations based on IoT and artificial intelligence anomaly detection accuracy and operational efficiency com-
technologies. Figures are usually the language that best pared to previous work. For example, temperature detection
connects academia with industry. In this sense, the proposed accuracy reached 98.7%, exceeding the reference values
system has shown considerable progress that deserves quan- by a notable margin. This result underlines the superiority
titative analysis. When comparing our system to conventional of our system in terms of reliability and effectiveness for
solutions, the most impactful metric is the consistent increase critical industrial applications. Furthermore, while other
in accuracy over time. Initially, our system operated with works have explored aspects of real-time monitoring and
an accuracy of 92% [47]. However, after implementing predictive analytics, very few have achieved the synergy
continuous feedback and adaptation processes, this figure between IoT and AI that our system offers. The quantitative
increased to an impressive 99% in just 12 weeks. It is rare analysis highlights our system’s unique ability to learn,
in contemporary literature to find a system that improves by adapt, and operate with exceptional precision under dynamic
7% quickly [49], [50]. In contrast, many systems referenced conditions [36], [54]. Our system works best in an industry
in previous research show marginal improvements of 1% where every percentage point can translate into significant
or 2% over similar periods; some even experience negative operational savings and competitive advantages. It sets a new
deviations due to a lack of adaptability [51], [52]. standard for future research and practical applications.
VOLUME 12, 2024 40383
 W. Villegas-Ch et al.: Toward Intelligent Monitoring in IoT: AI Applications

FIGURE 8. Evolution of performance with feedback.

According to the results obtained, it has been demonstrated continuous learning feature ensures that the system maintains
how our integrated IoT and AI system overcomes the its relevance over time and improves its ability to detect
limitations of conventional monitoring systems, offering a and predict anomalies. In an industrial environment that is
more efficient and adaptable solution for real-time predictive constantly evolving, this adaptability is invaluable.
analysis. Unlike existing approaches, our approach combines Comparing our solution with others on the market, it is
real-time data collection with advanced AI algorithms to evident that our proposal meets current demands and lays
predict failures before they occur, representing a significant the foundation for future advancements. Combining real-time
advancement in intelligent monitoring capability. monitoring with advanced predictive analytics is a direction
the global industry is moving in, and our solution is a
significant step in that direction.
V. CONCLUSION
While the system has proven its value in operational terms,
The need for advanced monitoring systems is paramount as
it is essential to consider the ethical implications of constant
the global industry moves towards an era of digitalization
monitoring. Ensuring that all data is collected and used
and enhanced connectivity. Our proposal, which skillfully
ethically and with due respect to privacy and autonomy is
combines the capabilities of IoT devices with advanced
crucial. As we continue to develop and refine our solution,
AI algorithms, has proven to be an effective and robust
ethical consideration of how this data is used and stored will
response to this demand. More than a simple data collection
be a top priority.
tool, this system has emerged as a comprehensive solution
There is ample room to improve and expand our
that promotes more proactive and predictive management of
proposition. With the constant evolution of AI technology,
industrial operations.
integrating more advanced algorithms can further improve
The direct impact of the system on daily operations has
the accuracy and speed of the system. Furthermore, with the
been immense. By providing early warnings about potential
growing trend towards Industry 4.0, incorporating more IoT
failures or anomalous trends, decisions are no longer made
devices and integration with automation systems can take the
only in response to problems once they have arisen. Instead,
solution to new horizons.
decisions are made proactively with foresight, avoiding
potentially costly failures or interruptions. This shift toward REFERENCES
more informed and proactive decision-making is crucial to [1] A. H. Sodhro, A. Gurtov, N. Zahid, S. Pirbhulal, L. Wang,
optimizing efficiency and reducing operating costs. M. M. U. Rahman, M. A. Imran, and Q. H. Abbasi, ‘‘Toward convergence
of AI and IoT for energy-efficient communication in smart homes,’’
One of the most notable aspects of the system is its IEEE Internet Things J., vol. 8, no. 12, pp. 9664–9671, Jun. 2021, doi:
ability to adapt and learn as it collects more data. This 10.1109/JIOT.2020.3023667.

40384 VOLUME 12, 2024

W. Villegas-Ch et al.: Toward Intelligent Monitoring in IoT: AI Applications

[2] S. M. Zayed, G. Attiya, A. El-Sayed, A. Sayed, and E. E.-D. Hemdan, [20] N. Tabassum and C. R. K. Reddyy, ‘‘Review on QoS and security
‘‘An efficient fault diagnosis framework for digital twins using optimized challenges associated with the Internet of Vehicles in cloud com-
machine learning models in smart industrial control systems,’’ Int. puting,’’ Meas., Sensors, vol. 27, Jun. 2023, Art. no. 100562, doi:
J. Comput. Intell. Syst., vol. 16, no. 1, pp. 2–18, May 2023, doi: 10.1016/j.measen.2022.100562.
10.1007/s44196-023-00241-6. [21] U. Ahmed, J. C.-W. Lin, and G. Srivastava, ‘‘A resource allocation deep
[3] V. K. Rathi, N. K. Rajput, S. Mishra, B. A. Grover, P. Tiwari, A. K. Jaiswal, active learning based on load balancer for network intrusion detection in
and M. S. Hossain, ‘‘An edge AI-enabled IoT healthcare monitoring system SDN sensors,’’ Comput. Commun., vol. 184, pp. 56–63, Feb. 2022, doi:
for smart cities,’’ Comput. Electr. Eng., vol. 96, Dec. 2021, Art. no. 107524, 10.1016/j.comcom.2021.12.009.
doi: 10.1016/j.compeleceng.2021.107524. [22] S. H. H. Nemati, N. Eslami, and M. H. Moaiyeri, ‘‘A hybrid SRAM/RRAM
[4] H. Atlam, R. Walters, and G. Wills, ‘‘Fog computing and the Internet in-memory computing architecture based on a reconfigurable SRAM
of Things: A review,’’ Big Data Cognit. Comput., vol. 2, no. 2, p. 10, sense amplifier,’’ IEEE Access, vol. 11, pp. 72159–72171, 2023, doi:
Apr. 2018, doi: 10.3390/bdcc2020010. 10.1109/ACCESS.2023.3294675.
[5] M. Ding, ‘‘Research on tourism route planning based on artificial [23] A. Taïk, Z. Mlika, and S. Cherkaoui, ‘‘Data-aware device scheduling for
intelligence technology,’’ Wireless Commun. Mobile Comput., vol. 2021, federated edge learning,’’ IEEE Trans. Cognit. Commun. Netw., vol. 8,
pp. 1–7, Dec. 2021, doi: 10.1155/2021/2227798. no. 1, pp. 408–421, Mar. 2022, doi: 10.1109/TCCN.2021.3100574.
[6] F. Alshehri and G. Muhammad, ‘‘A comprehensive survey of the Internet [24] T. Saba, A. Rehman, K. Haseeb, T. Alam, and G. Jeon, ‘‘Cloud-
of Things (IoT) and AI-based smart healthcare,’’ IEEE Access, vol. 9, edge load balancing distributed protocol for IoE services using swarm
pp. 3660–3678, 2021, doi: 10.1109/ACCESS.2020.3047960. intelligence,’’ Cluster Comput., vol. 26, no. 5, pp. 2921–2931, Oct. 2023,
[7] R. Calegari, E. Denti, S. Mariani, and A. Omicini, ‘‘Logic programming doi: 10.1007/s10586-022-03916-5.
as a service in multi-agent systems for the Internet of Things,’’ Int. J. [25] R. Mehta, V. K. Prasad, S. Mishra, S. Tanwar, and Y. Patel, ‘‘Evolving
Grid Utility Comput., vol. 10, no. 4, 2019, doi: 10.1504/IJGUC.2019. technologies: IoT and artificial intelligence for healthcare informatics,’’
10022135. in Innovations in Healthcare Informatics: From Interoperability to Data
[8] A. Alti, A. Lakehal, S. Laborie, and P. Roose, ‘‘Autonomic semantic- Analysis, M. Abouhawwash, S. Tanwar, A. Nayyar, and M. Naved,
based context-aware platform for mobile applications in pervasive Eds. U.K.: The Institution of Engineering and Technology, 2023, doi:
environments,’’ Future Internet, vol. 8, no. 4, p. 48, Sep. 2016, doi: 10.1049/pbhe041e_ch8.
10.3390/fi8040048. [26] O. Hadj Abdelkader, H. Bouzebiba, D. Pena, and A. P. Aguiar, ‘‘Energy-
[9] N. Fredj, Y. H. Kacem, S. Khriji, O. Kanoun, and M. Abid, ‘‘A efficient IoT-based light control system in smart indoor agriculture,’’
review on intelligent IoT systems design methodologies,’’ Meas., Sen- Sensors, vol. 23, no. 18, p. 7670, Sep. 2023, doi: 10.3390/s23187670.
sors, vol. 18, Dec. 2021, Art. no. 100347, doi: 10.1016/j.measen.2021. [27] A. B. Kathole, K. N. Vhatkar, S. Kumbhare, J. Katti, and V. V. Kimbahune,
100347. ‘‘IoT-based smart agriculture for onion plant disease management:
[10] V. Kelli, V. Argyriou, T. Lagkas, G. Fragulis, E. Grigoriou, and A comprehensive approach,’’ Int. J. Intell. Syst. Appl. Eng., vol. 12, no. 13,
P. Sarigiannidis, ‘‘IDS for industrial applications: A federated learning pp. 472–476, 2024.
approach with active personalization,’’ Sensors, vol. 21, no. 20, p. 6743,
[28] A. B. Kathole, K. N. Vhatkar, and S. D. Patil, ‘‘IoT-enabled pest
Oct. 2021, doi: 10.3390/s21206743.
identification and classification with new meta-heuristic-based deep
[11] M. Murugan, J. Rajarajan, G. Hariharan, and S. Jeevitha, ‘‘Industry
learning framework,’’ Cybern. Syst., vol. 55, no. 2, pp. 380–408, Feb. 2024,
4.0: A real-time implementation of automation and digitalization in an
doi: 10.1080/01969722.2022.2122001.
industry,’’ Natural Volatiles Essential Oils, vol. 8, no. 4, pp. 1568–1580,
[29] S. Pitafi, T. Anwar, I. D. M. Widia, and B. Yimwadsana, ‘‘Revo-
Nov. 2021.
lutionizing perimeter intrusion detection: A machine learning-driven
[12] P. Baum-Talmor and M. Kitada, ‘‘Industry 4.0 in shipping: Implications to
approach with curated dataset generation for enhanced security,’’ IEEE
seafarers’ skills and training,’’ Transp. Res. Interdiscipl. Perspect., vol. 13,
Access, vol. 11, pp. 106954–106966, 2023, doi: 10.1109/ACCESS.2023.
Mar. 2022, Art. no. 100542, doi: 10.1016/j.trip.2022.100542.
3318600.
[13] O. Bongomin, G. Gilibrays Ocen, E. Oyondi Nganyi, A. Musinguzi,
[30] T. Serrano-Ramírez, N. D. C. Lozano-Rincón, A. Mandujano-Nava,
and T. Omara, ‘‘Exponential disruptive technologies and the required
and Y. J. Sámano-Flores, ‘‘Artificial vision system for object classi-
skills of Industry 4.0,’’ J. Eng., vol. 2020, pp. 1–17, Feb. 2020, doi:
fication in real time using raspberry pi and a web camera,’’ Revista
10.1155/2020/4280156.
Tecnologías Información Comunicaciones, vol. 5, pp. 20–25, Jun. 2021,
[14] G. Filios, S. Karatzas, M. Krousarlis, S. Nikoletseas, S. H. Panagiotou,
doi: 10.35429/jitc.2021.13.5.20.25.
and P. Spirakis, ‘‘Realizing predictive maintenance in production
machinery through low-cost IIoT framework and anomaly detection: [31] P. Lemenkova, ‘‘Processing oceanographic data by Python libraries numpy,
A case study in a real-world manufacturing environment,’’ in Proc. scipy and pandas,’’ Aquatic Res., vol. 2, no. 2, pp. 73–91, 2019.
19th Int. Conf. Distrib. Comput. Smart Syst. Internet Things (DCOSS- [32] C. L. Vidal-Silva, A. Sánchez-Ortiz, J. Serrano, and J. M. Rubio,
IoT), Jun. 2023, pp. 338–346, doi: 10.1109/DCOSS-IoT58021.2023. ‘‘Academic experience in rapid development of web information systems
00062. with Python and Django,’’ Formación Universitaria, vol. 14, no. 5, p. 85,
[15] V. Sima, I. G. Gheorghe, J. Subić, and D. Nancu, ‘‘Influences of the 2021.
Industry 4.0 revolution on the human capital development and consumer [33] C. Ewuoso and S. Hall, ‘‘Core aspects of ubuntu: A systematic review,’’
behavior: A systematic review,’’ Sustainability, vol. 12, no. 10, p. 4035, South Afr. J. Bioethics Law, vol. 12, no. 2, p. 93, Dec. 2019, doi:
May 2020, doi: 10.3390/su12104035. 10.7196/sajbl.2019.v12i2.679.
[16] W. Jiang, ‘‘A machine vision anomaly detection system to Industry 4.0 [34] F. Zhu, W. Cai, Z. Wang, and F. Li, ‘‘AI-empowered propagation
based on variational fuzzy autoencoder,’’ Comput. Intell. Neurosci., prediction and optimization for reconfigurable wireless networks,’’
vol. 2022, pp. 1–10, Mar. 2022, doi: 10.1155/2022/1945507. Wireless Commun. Mobile Comput., vol. 2022, pp. 1–10, Jan. 2022, doi:
[17] B. Wachnik, ‘‘Analysis of the use of artificial intelligence in the 10.1155/2022/9901960.
management of Industry 4.0 projects. The perspective of Polish indus- [35] K. Dheyaa Ismael and S. Irina, ‘‘Face recognition using Viola–Jones
try,’’ Prod. Eng. Arch., vol. 28, no. 1, pp. 56–63, Mar. 2022, doi: depending on Python,’’ Indonesian J. Electr. Eng. Comput. Sci., vol. 20,
10.30657/pea.2022.28.07. no. 3, p. 1513, Dec. 2020, doi: 10.11591/ijeecs.v20.i3.pp1513-1521.
[18] F. Martell, J. M. López, I. Y. Sánchez, C. A. Paredes, and E. Pisano, [36] B. Pardamean, T. Suparyanto, A. Anugrahana, I. Anugraheni, and
‘‘Evaluation of the degree of automation and digitalization using a D. Sudigyo, ‘‘Implementasi team-based learning dalam pengembangan
diagnostic and analysis tool for a methodological implementation of pembelajaran online berbasis artificial intelligence,’’ Scholaria, Jurnal
Industry 4.0,’’ Comput. Ind. Eng., vol. 177, Mar. 2023, Art. no. 109097, Pendidikan Kebudayaan, vol. 12, no. 2, pp. 118–126, May 2022, doi:
doi: 10.1016/j.cie.2023.109097. 10.24246/j.js.2022.v12.i2.p118-126.
[19] S. B. Atitallah, M. Driss, W. Boulila, and H. B. Ghézala, ‘‘Lever- [37] X. Liu, W. Zhang, X. Tong, F. Zhong, Z. Li, Z. Xiong, J. Xiong, X. Wu,
aging deep learning and IoT big data analytics to support the Z. Fu, X. Tan, Z. Liu, S. Zhang, H. Jiang, X. Li, and M. Zheng, ‘‘MolFil-
smart cities development: Review and future directions,’’ Comput. Sci. terGAN: A progressively augmented generative adversarial network for
Rev., vol. 38, Nov. 2020, Art. no. 100303, doi: 10.1016/j.cosrev.2020. triaging AI-designed molecules,’’ J. Cheminform., vol. 15, no. 1, pp. 2–14,
100303. Apr. 2023, doi: 10.1186/s13321-023-00711-1.

VOLUME 12, 2024 40385

 W. Villegas-Ch et al.: Toward Intelligent Monitoring in IoT: AI Applications

[38] Y. W. Oh and C. H. Park, ‘‘User perceptions of AI-based WILLIAM VILLEGAS-CH received the master’s
comment filtering technology,’’ Amer. Behav. Scientist, May 2023, degree in communications networks and the Ph.D.
Art. no. 000276422311743, doi: 10.1177/00027642231174331. degree in computer science from the University of
[39] A. A. A. Alshawi, J. Tanha, and M. A. Balafar, ‘‘An attention- Alicante. He is currently a Professor of informa-
based convolutional recurrent neural networks for scene text tion technology with Universidad de Las Améri-
recognition,’’ IEEE Access, vol. 12, pp. 8123–8134, 2024, doi: cas, Quito, Ecuador. He is a Systems Engineer
10.1109/ACCESS.2024.3352748. specializing in robotics in artificial intelligence.
[40] F. Luna-Perejón, M. J. Domínguez-Morales, and A. Civit-Balcells,
He has participated in various conferences as a
‘‘Wearable fall detector using recurrent neural networks,’’ Sensors, vol. 19,
Speaker on topics, such as ICT in education and
no. 22, p. 4885, Nov. 2019, doi: 10.3390/s19224885.
[41] A. Krizhevsky, I. Sutskever, and G. E. Hinton, ‘‘ImageNet classification
how they improve educational quality and student
with deep convolutional neural networks,’’ Commun. ACM, vol. 60, no. 6, learning. His main articles focus on the design of ICT systems, models,
pp. 84–90, May 2017, doi: 10.1145/3065386. and prototypes applied to different academic environments, especially with
[42] B. Liang, H. Li, M. Su, X. Li, W. Shi, and X. Wang, ‘‘Detecting adversarial the use of big data and artificial intelligence as a basis for creating
image examples in deep neural networks with adaptive noise reduction,’’ intelligent educational environments. His main research interests include
IEEE Trans. Dependable Secure Comput., vol. 18, no. 1, pp. 72–85, web applications, data mining, and e-learning.
Jan. 2021, doi: 10.1109/TDSC.2018.2874243.
[43] A. Ramesh, C. K. Subbaraya, and G. K. R. Kumar, ‘‘AI based
dynamic prediction model for mobile health application system,’’ Int.
J. Adv. Comput. Sci. Appl., vol. 14, no. 1, pp. 356–365, 2023, doi:
10.14569/ijacsa.2023.0140138.
[44] D. C. Niehorster, L. Li, and M. Lappe, ‘‘The accuracy and precision
of position and orientation tracking in the HTC vive virtual reality
system for scientific research,’’ I-Perception, vol. 8, no. 3, Jun. 2017,
Art. no. 204166951770820, doi: 10.1177/2041669517708205.
[45] A. P. Hunt, A. J. E. Bach, D. N. Borg, J. T. Costello, and I. B. Stewart, ‘‘The
systematic bias of ingestible core temperature sensors requires a correction
by linear regression,’’ Frontiers Physiol., vol. 8, pp. 1–7, Apr. 2017, doi:
10.3389/fphys.2017.00260. JOSELIN GARCÍA-ORTIZ is currently pursuing
[46] J. Deutscher, K. Gutjahr, R. Perko, H. Raggam, M. Hirschmugl, and the degree in cybersecurity engineering with
M. Schardt, ‘‘Humid tropical forest monitoring with multi-temporal L-, C- Universidad de Las Américas. She is a Software
and X-band SAR data,’’ in Proc. 9th Int. Workshop Anal. Multitemporal
Engineer with Capmation Inc., Quito, Ecuador,
Remote Sens. Images (MultiTemp), Jun. 2017, pp. 1–4, doi: 10.1109/Multi-
where she applies her knowledge in software
Temp.2017.8035264.
development and cybersecurity. She has a strong
[47] B. A. Lynch, T. K. Kaufman, T. I. Rajjo, K. Mohammed, S. Kumar,
M. H. Murad, N. E. Gentile, G. A. Koepp, S. K. McCrady-Spitzer, interest in the intersection of AI and its applica-
and J. A. Levine, ‘‘Accuracy of accelerometers for measuring physical tions to education and has previous experience in
activity and levels of sedentary behavior in children: A systematic researching the subject.
review,’’ J. Primary Care Community Health, vol. 10, Jan. 2019,
Art. no. 215013271987425, doi: 10.1177/2150132719874252.
[48] A. Sivanathan, H. H. Gharakheili, and V. Sivaraman, ‘‘Detecting
behavioral change of IoT devices using clustering-based network traffic
modeling,’’ IEEE Internet Things J., vol. 7, no. 8, pp. 7295–7309,
Aug. 2020, doi: 10.1109/JIOT.2020.2984030.
[49] J. Alvarez-Montoya, A. Carvajal-Castrillón, and J. Sierra-Pérez, ‘‘In-flight
and wireless damage detection in a UAV composite wing using fiber optic
sensors and strain field pattern recognition,’’ Mech. Syst. Signal Process.,
vol. 136, Feb. 2020, Art. no. 106526, doi: 10.1016/j.ymssp.2019.106526.
[50] J. B. Carter, R. Huffaker, A. Singh, and E. Bean, ‘‘HUM: A review of
hydrochemical analysis using ultraviolet-visible absorption spectroscopy
and machine learning,’’ Sci. Total Environ., vol. 901, Nov. 2023,
Art. no. 165826, doi: 10.1016/j.scitotenv.2023.165826. SANTIAGO SÁNCHEZ-VITERI was born in
[51] S. Shohan, Y. Zeng, X. Chen, R. Jin, and R. Shirwaiker, ‘‘Investigating Quito, Ecuador, in 1984. He received the Systems
dielectric spectroscopy and soft sensing for nondestructive quality Engineering degree from Salesian Polytechnic
assessment of engineered tissues,’’ Biosensors Bioelectron., vol. 216,
University (UPS), Ecuador, in 2017, where he
Nov. 2022, Art. no. 114286, doi: 10.1016/j.bios.2022.114286.
is currently pursuing the master’s degree in
[52] S. Ghosh, S. Kim, M. F. Ijaz, P. K. Singh, and M. Mahmud, ‘‘Classification
telematics management. He work in the area
of mental stress from wearable physiological sensors using image-
encoding-based deep neural network,’’ Biosensors, vol. 12, no. 12, p. 1153, of telecommunications and networking of IT
Dec. 2022, doi: 10.3390/bios12121153. with Universidad Internacional del Ecuador, since
[53] L. Zhao, S. J. Pan, and Q. Yang, ‘‘A unified framework of active 2010. He has been a Professor of computer science
transfer learning for cross-system recommendation,’’ Artif. Intell., vol. 245, with Universidad Internacional del Ecuador. He is
pp. 38–55, Apr. 2017, doi: 10.1016/j.artint.2016.12.004. an Administrator of computer servers and internet connection equipment.
[54] L. Huo, D. Jiang, S. Qi, H. Song, and L. Miao, ‘‘An AI-based adaptive He is an Administrator of Moodle and CANVAS learning management
cognitive modeling and measurement method of network traffic for system with Universidad Internacional del Ecuador. He participated in more
EIS,’’ Mobile Netw. Appl., vol. 26, no. 2, pp. 575–585, Apr. 2021, doi: than ten indexed articles on telecommunications, big data, and innovation.
10.1007/s11036-019-01419-z.

40386 VOLUME 12, 2024