IoT Platform to monitor Industrial Machines and

perform PdM at the edge: A case study

Lahis Almeida*, Erimar Farias*, Vicente Farias*, Camilo Sidou*, Brunna
Talita*, Rafaela Lima*, Suelem Valadão‡, Lívia Araújo*
*R&D Projects Department, INDT - Instituto de Desenvolvimento Tecnológico, Manaus, Brazil
‡R&D Technical coordination of projects, ADATA Eletronics, Manaus, Brazil
{lahis.almeida, erimar.farias, vicente.farias, camilo.sidou, brunna.talita, rafaela.lima, livia.araujo}@indt.org.br
suelem valadao@adata.com

Abstract— Predictive maintenance plays a crucial role in In PdM applications, the challenges of handling big data
Industry 4.0 ensuring efficiency, product quality and a safe are magnified by the need for real time monitoring and
working environment. Edge computing is another trend that processing of a significant volume of data generated by
optimizes Smart Factories applications, bringing processing and sensors [1]. This poses difficulties in ensuring satisfactory
decision making closer to the data collection end node points, metrics such as latency, scalability, and network bandwidth,
reducing network traffic and latency. This paper proposed an particularly when immediate action is required to prevent
IoT platform to monitor industrial machines and perform failures caused by predicted events [4]. To address this,
predictive maintenance at the network edge. The effectiveness of another significant trend emerges within the Industry 4.0
the proposed platform is validated in a real industrial application
paradigm: Edge computing. By bringing processing closer to
project. Among the models considered in this study, the Decision
Tree model emerged as the most effective (reaching 96% of
the data collection point and delegating non-time-sensitive
accuracy) for the industrial machines analyzed. tasks to the cloud, edge computing effectively reduces the
strain on the network [5]. This approach optimizes the
Keywords—predictive maintenance, pdm, machine learning, distribution of processing workloads across edge nodes,
edge computing, industry 4.0, esp32 board. alleviating core network traffic and enhancing overall
application performance [6].
Within all this context of Industry 4.0 and smart
I. INTRODUCTION applications, the company ADATA Eletronics¹ (Amazonas -
Manaus' site), unfortunately did not have an automated and
This Industry 4.0 revolutionizes traditional industrial
integrated control of its facilities equipment (e.g., Generator,
automation by introducing advancements that lead to the
Compressor, QTA, Trafo), leading to some unwanted losses
creation of smart factories [1]. These facilities incorporate
and assets costs. In contrast to this scenario, the Utility
emergent technologies such as advanced robotics, high
Automation Research and Development (R&D) project was
computing power and connectivity [2]. Notably, remarkable
created, which consists of a partnership between INDT
advancements in the fields of Artificial Intelligence (AI) and
(Instituto de Desenvolvimento Tecnológico) and ADATA team,
Internet of Things (IoT) are fueling unprecedented
whose objective is to improve ADATA's maintenance process
technological and scientific developments [1]. By harnessing
through the development of a monitoring and control platform
the potential of Industry 4.0, smart factories are able to offer
for industrial utility equipment (i.e., Generator, UPS, QTA,
customers a range of new services and products that exhibit
Chiller). Thus, the main contribution of this work is to (i)
higher levels of efficiency and quality standards compared to
propose an IoT platform for monitoring ADATA industrial
previous approaches [2].
machines and perform PdM at the edge of the network; (ii)
Predictive maintenance (PdM) stands out as one of the testing and validate the proposed platform together with the
most prominent opportunities discussed in the realm of client; (iii) presents all steps to performs PdM at the edge and
Industry 4.0 [2]. It relies on ongoing monitoring of machine or (iv) evaluates PdM models through machine learning and
process integrity, enabling maintenance to be carried out performance metrics.
solely when necessary. Furthermore, PdM facilitates early
The rest of the paper is organized as follows: Section II
detection of failures through the application of predictive tools
provides an overview of related work that involve PdM
that leverage historical data (i.e., machine learning
application at the edge, Section III describes the proposed
techniques), statistical inference methods and engineering
approach employed in our research, Section IV presents the
strategies[3]. In an industrial environment, the implementation
platform experiments, deploy and the discussion of the results
of PdM applications has the potential to optimize production
obtained; and, finally, Section VI presents the conclusions and
profits while minimizing costs and losses, including those
future works directions.
associated with assets [4].

¹https://www.adata.com/br

ISSN 2447-8326. V.1 © 2023 BTSYM

 2023 Brazilian Technology Symposium

Fig. 1. Utility Project architecture. However, results showed that the error correction and
compensation process is mainly dependent on the types of
machines and processes. In [12] work, a total of six distinct
II. RELATED WORK supervised learning techniques were utilized for the PdM
approach. These methods encompass k-Nearest Neighbour
The continuous advancement of the AI and IoT (kNN), Back-propagation Feed-forward Neural Network
technologies has garnered significant attention in the research (FFNN), Decision Tree (DT), and Naïve Bayesian. The
community towards the PdM field [7]. In [8], the authors outcomes obtained from these approaches demonstrated
found that an IoT PdM system had the potential to enhance significant advantages in terms of economic benefits when
production throughput, minimize maintenance costs, and compared to conventional reactive maintenance and
facilitate effective decision-making. Their research showed time-based maintenance practices. The findings revealed an
that several design considerations are taken into account when estimated cost reduction of 40%, coupled with a remarkable
this kind of IoT system is proposed, including the types of six-fold decrease in the number of breakdowns.
analysis available, communication technology options, and the
sensors required for data collection. In this PdM context, the present work proposes to
contribute with insights and discussions about the PdM
In [5] study, a data driven PdM system was developed for solutions through an IoT platform to monitor industrial
production lines in manufacturing. They utilize real-time data machines and perform PdM at the edge of the network.
from IoT sensors and employ machine learning methods to Additional contributions concern in showing all challenges
identify signals indicative of potential failures. The evaluation and steps to perform PdM at constrained devices (32 bit
results demonstrate the effectiveness of the proposed PdM architecture) in a real world industrial application.
system and indicate that boosting and bagging ensemble
models perform well when the output of the model represents III. PROPOSED APPROACH
the estimated useful time remaining before a failure.
This paper proposes a methodology for monitoring
In [9], the authors proposed a PdM architecture based on electrical equipment in industrial environments, enabling end
containerized micro services on intelligent edge devices users, in this case, ADATA's facilities team, to remotely
together with a hybrid model which fuses generalized fault monitor the behavior of this equipment, identifying anomalies
trees (GFTs) and anomaly detection. Their results revealed a present in them. In addition to monitoring the variables of
significant estimated reduction of over 63% in current these devices (e.g. voltage, current, power factor), a PdM
maintenance costs. In [10], authors focused on estimating the module is also proposed at the edge, i.e., at the end nodes, in
remaining useful life (RUL) of electronic components through order to mitigate network latency and enable faster decision
a three-step process: selecting fault features using variance making. The following sections describe the monitoring
analysis, measuring fault severity with relative entropy architecture and the project architecture is presented in Fig. 1.
distance, and conducting fault prognosis using the gradient
The first stage of the proposed approach consists in
boosting decision tree (GBDT) model. Their results
establishing communication between the industrial machine
demonstrated superior prediction accuracy compared to other
(I) and the end node (II). The industrial machine component
distance-measuring methods for determining the RUL of
represents the equipment that will be monitored. In ADATA,
electronic elements.
there are equipments located in outdoor and indoor
The literature covers a wide variety of ML techniques for environments.
PdM (i.e., supervised classification, anomaly detection,
The end node (II) component of the Utility Control project
regression in high dimensional data), each with specific
is the component responsible for wired communication with
characteristics and applications [1][7]. The authors of [11]
the machines and inductive sensors installed in the ADATA
used ANN predictive maintenance in machine centers.
plant; and for wireless communication with the Dashboard

ISSN 2447-8326. V.1 © 2023 BTSYM

 2023 Brazilian Technology Symposium
Server (IV) through the WiFi/WiFi Mesh Network (III). It is
also responsible for inferring the PdM models in the respective
selected equipment, sending the respective alert notifications
if any anomalous behavior is identified.
Most equipment such as QTA, UPS, Generator,
Compressor, have HMIs (Human Machine Interface) with
communication interface via MODBUS RTU, enabling
communication with the end node of the solution. However,
there are some devices that do not have this interface such as
the Diesel Tank, Exhausters and Insulators; therefore, for these
cases, isolated sensors were installed and communication with
them was also done through the end node, by protocols (e.g.,
I2C, 4-20mA).
Fig. 2. End node.
After the connection established between industrial
machine and end node, the collection of the variables of Finally, the firmware also had the PdM module, which is
interest of the machine is carried out and the data are responsible for performing the inferences of the PdM models
transmitted through the WiFi Mesh network (III) to a central in real time. To manage all firmware modules (i.e., sensor
end node, which follows the transmission, via traditional WiFi interfaces, HTTPS communication, PdM inferences), the code
and the HTTPs protocol, sending them to the Server (IV) and was built based in FreeRTOS³ tasks approach.
the developed Dashboard software (V), This last component
aims to present the collected variables in various visualization B. Dashboard
formats (e.g., widgets, pop-ups, graphs) for better monitoring The implementation of the Dashboard software screens
of the facilities team. In this work, WiFi Mesh network will was done using the AngularJS web framework, with the
not be described in detail given the complexity of its Angular Material UI library to implement most of the screen
arrangement. However, in future opportunities, mesh network components, following the Material Design standard. For the
coverage results can be detailed and disseminated. development of the API, the ASP.NET Core framework was
chosen, following the REST standards for web application
IV. EXPERIMENTS AND RESULTS
development. The choice of ASP.NET Core is due to some of
The experiments of this work aimed to validate the the existing systems in the ADATA company having been
proposed IoT monitoring platform as well as to evaluate the previously developed in the same language, C#, in addition to
predictive maintenance AI module at the edge. The following being a very robust web application development platform,
subsections describe the experiments and obtained results. offering a series of tools and development libraries that
facilitate the delivery of functionalities [13].
A. End node
The end node PCB design of the Utility Automation
project was developed using Altium Designer² software in
version 22.1.2. The System on a Chip (SoC) used in the
project was the ESP32-WROOM-32UE, manufactured by
Espressif, due its main features are the support for WiFi
2.4GHz, 32 bit microprocessor with two cores (Xtensa), 26
GPIOs and connector for external antenna, thus meeting the
hardware requirements of the our project. The main circuits
designed for the end node were: recording circuit (USB/TTL),
interface with digital (e.g., temperature, water flow) and
analog sensors (e.g., current, pressure) and the communication
interfaces (i.e, I2C, RS485 (MODBUS RTU), ADC). The
fabricated and assembled PCB can be visualized in Fig. 2.
The implementation of the firmware modules in the end
node was done using the official Espressif framework,
ESP-IDF (version 4.4). The languages used for
implementation were C/C++. In addition to the Fig. 3. UPS monitoring screen.
communication modules with the sensors and machines,
WiFi/WiFi Mesh communication was implemented and, in the The open source relational PostgreSQL system was chosen
application layer, the HTTPS protocol was implemented. In as a database due extensive documentation and support on
the firmware, a data persistence module was also different platforms and libraries [14]. For integration between
implemented, using SPIFFS (SPI Flash File System), so that the web application and its database was used the Entity
even in an eventual reset caused by an energy peak, the board Framework Core, which is a native integration solution for
does not lose registration information such as ID and web applications developed in ASP.NET Core. Fig. 3 shows
commands coming from the server.

ISSN 2447-8326. V.1 © 2023 BTSYM

 2023 Brazilian Technology Symposium
one of the dashboard screens, presenting one of the monitoring
screens (UPS machine Dashboard).
C. Predictive maintenance at the edge

After analysis in the literature (Section II) and alignments

with the client, it was decided to treat the PdM module of the
Utility Automation project as a type of binary classification
problem (Supervised Learning), that is, the output of the
model would be 0, characterizing a normal behavior in the
respective machine (not anomalous); or 1, characterizing an
abnormal problem in the machine (anomalous).

The predictive maintenance models embedded in the end

nodes will perform anomaly or non-anomaly inferences
according to the information collected from the respective
equipment selected to have the PdM module. Thus, for each
UPS type equipment, i.e., the model that was trained with the
UPS equipment information will be embedded in the end node
for monitoring this equipment. The steps for the construction
of the PdM module are:

● Definition of the target machines of the PdM module;

● Data collection;
Fig. 4. UPS machine.
● Data preprocessing;
● Training of the models; failures in this equipment. The monitoring logs of the
● Deployment of models in the end nodes; equipment were also not accessible and the bureaucracy
between ADATA and the manufacturers of these equipment
● Evaluation of PdM models. was hampered by the costs of technical visits and availability
of the technicians outsourced by them; making it impossible to
Following subsections detail the completion of these steps.
use the logs as dataset in our project without causing losses in
1) PDM module target machines: the equipment chosen delivery time.
to have the PdM module were the QTA and UPS. The QTA In this context, data field collection was needed. The data
(Automatic Transfer Board) equipment is responsible for collection performed through MODBUS RTU communication
triggering the start of equipment (e.g., generator sets) between the endnode and the respective equipment lasted
immediately after the interruption/failure of the Energy approximately 1 month. This period was defined according to
Provider's power, without the manual exposure of employees. the planning and delivery deadlines of the Utility Automation
Thus, in situations where the main power, i.e. the power project. After this period, the collected data were exported
supplying ADATA, is interrupted, the QTA automatically through the Dashboard software itself in the csv format. For
triggers the company's internal generators so that no factory the QTA equipment, the variables collected were the three
environment is preadjudicated [15]. UPS (Uninterruptible phases of the input voltage (V), three phases of the input
Power Supply) equipment is used to protect data centers and current (A), power factor, frequency (Hz) and instantaneous
electrical equipment that have a high level of criticality, such consumption (kWh), while for the UPS equipment, the
as those operating on production lines. The main function of variables collected were the three phases of the output voltage
the UPS is to supply power to critical loads when the main (V) and three phases of the output current (A). Data collection
power fails. The UPS differs from the QTA in that its response frequency for both equipment was 20Hz. For QTA, a total of
time is instantaneous, as it delivers energy stored in its internal 197.337 samples were purchased, and for UPS, 196.453.
batteries, and is sufficiently designed to withstand periods of 3) Data preprocessing: in this phase, the raw data
time until the main power is switched to a secondary power collected undergo treatments that allow them to be more
source (Generator) [16]. In this context, it is possible to robust for analysis. The first treatment employed on the data
observe the great relevance of these two functions for was the separation of samples into binary classes: abnormal
ADATA, because without them, an eventual power failure and normal. For samples that presented values within the
could cause great losses in its manufacturing processes and expected ranges, the data were labeled as normal. Otherwise,
industrial tests. Fig. 3 shows the UPS industrial equipment. they were labeled as abnormal. After the separation into
2) Data collection: the implementation of the PdM classes, the missing data check was performed, i.e., if there
module at ADATA faced a number of challenges since its was missing data between the sensor readings of the collected
conception. ADATA's site in Manaus is recent (~ 2 years) and, samples; and if the classes were balanced. After the missing
since the installation of the industrial machines (e.g., data check, it was possible to observe that there was no
Generator, QTA, Compressor, UPS) there was no history of occurrence of missing data. However, the two datasets were

ISSN 2447-8326. V.1 © 2023 BTSYM

 2023 Brazilian Technology Symposium
unbalanced. The QTA dataset had 189.455 of normal behavior others, besides its port process to a cpp lib was more complex
and 7.882 of abnormal behavior. The UPS dataset had 196.322 due code size generated. Thus, for QTA equipment it was also
of normal behavior and 131 of abnormal behavior. To mitigate chosen to be the Decision Tree model to be used in the PdM
this problem, the oversampling technique (90%) was module.
performed on the data. The anomaly data that were generated
were based on the maximum and minimum ranges of each of TABLE II. QTA PDM COMPRESSED MODELS RESULTS
the variables collected in the samples. As a consequence, the QTA
abnormal class of the QTA dataset became 189.430 and the Mean
Compressed
UPS, 196.586. Algorithm Accuracy
Model Size
inference
4) Training of the models: The PdM models were time
trained on a Dell Vostro 15 model 5510 notebook, with the Naive Bayes 84 % 1.6 kB 10 us
11th Generation Intel Core I7-11390H processor with a Decision Tree 96 % 1.9 kB 13 us
frequency of 3.4GHz; and 16GB memory and 512GB SSD.
The algorithms chosen to train the models were Decision Tree, Random Forest 96 % 3.4 kB -
Random Forest and Naive Bayes as they are widely used in
PdM classification problem perspective. Once these steps were completed, the IoT project to
5) Deployment of the respective models in the end monitor Industrial Machines and perform Predictive
nodes: in this step, PdM models are ported to 32-bit Maintenance at the edge was implemented at ADATA, which
architecture (Xtensa chip). The models are compressed into a led to the automation and integration of Adata's industrial
.h file, using the MicroML generator4 library, and from there it equipment. ADATA's facilities team validated the developed
is possible to insert it into the firmware project and perform software, as well as the installed sensors and end nodes.
the model inference calls. To perform the inference correctly, Regarding the PdM module some occurrences have already
the same vector of variables per sample (feature vector) been identified. All related to power surges or any possible
considered in the training must be used, so, for UPS, for machine shutdown, proving that the deployed model can
example, the input vector must always be the three phases of detect anomalous behaviors and notify, through alarm
current and output voltage of the equipment. Table 1 shows a notifications on the developed Dashboard. Fig. 5 shows the
comparison between the size of the compressed model and the anomaly notification screen.
characteristics of its operation in the end node.
6) Evaluation of PdM models: the metric chosen to
evaluate the classification algorithms was Accuracy, as it is
responsible for measuring the percentage of correct
predictions that an AI model made [17]. Regarding the
performance of the model on the end node, some interesting
features were listed such as compressed model size and
inference time. All these evaluated methods and data insights
can assist to choose the most indicated model to be deployed
in which end node.

Evaluating Table I and II, the algorithm that had the best
performance on UPS end node was Decision Tree, since its
accuracy result (98 %) was superior to the others. Another
differential of this model, was inference time that leads to a
better latency and very compact model size, facilitating its Fig. 5. UPS detected an anomaly alarm screen.
execution on firmware project.
V. CONCLUSION
TABLE I. UPS PDM COMPRESSED MODELS RESULTS
In this research, an IoT platform was developed to monitor
UPS and control the equipment of ADATA's industrial plant,
Mean improving their maintenance process. In addition, a PdM
Compressed
Algorithm Accuracy inference
Model Size
time module on the edge was also developed, helping in the
Naive Bayes 87 % 2.4 kB 34 us
decision making of maintenance schedules.

Decision Tree 98 % 1.2 kB 6 us The results of the research were successful as they showed
that the proposed flow approach for the IoT platform worked
Random Forest 93 % 228 kB - well in the industrial environment considered. The INDT and
ADATA team conducted thorough testing of the end node and
For the QTA end node, Decision Tree and Random Dashboard components, validating the widget views and
Forest models were the best results (both with 96%). CRUD operations in the developed software.
Although, Random Forest compressed model are larger than

ISSN 2447-8326. V.1 © 2023 BTSYM

 2023 Brazilian Technology Symposium
Despite the challenges faced in the implementation of the [6] “Predictive maintenance and the smart factory,”
https://www2.deloitte.com/content/dam/Deloitte/us/Documents/process-
AI module, such as difficulty in obtaining the equipment
and-operations/us-cons-predictive-maintenance.pdf, 2023, accessed:
operation logs for use as dataset; it was possible to trace an 2023-07-17.
effective flow of steps from data collection in the field to the [7] S. Ayvaz and K. Alpay, “Predictive maintenance system for production
porting of the models to the end nodes of the selected lines in manufacturing: A machine learning approach using iot data in
equipment (QTA and UPS). The two trained models showed real-time,” Expert Systems with Applications, vol. 173, p. 114598, 2021.
accuracy above 96% and the adopted flow can be used as a [8] B. Sicard, N. Alsadi, P. Spachos, Y. Ziada, and S. A. Gadsden,
basis for more equipment to have PdM module in future “Predictive maintenance and condition monitoring in machine tools: An
iot approach,” in 2022 IEEE International IOT, Electronics and
works. Thus, we hope that our research will contribute to Mechatronics Conference (IEMTRONICS). IEEE, 2022, pp. 1–9.
discussions and insights around PdM and Edge computing [9] P. Nunes, E. Rocha, and J. P. Santos, “Using intelligent edge devices for
areas. predictive maintenance on injection molds,” Applied Sciences, vol. 13,
no. 12, p. 7131, 2023.
Future opportunities for the PdM module include: dealing [10] L. Wang, D. Zhou, H. Zhang, W. Zhang, and J. Chen, “Application of
with the PdM problem from the Regression perspective, as the relative entropy and gradient boosting decision tree to fault prognosis in
Remain Time Useful information may also be interesting for electronic circuits,” Symmetry, vol. 10, no. 10, p. 495, 2018.
the facilities team. In the data pre-processing stage, employ [11] Z. Li, Y. Wang, and K.-S. Wang, “Intelligent predictive maintenance for
fault diagnosis and prognosis in machine centers: Industry 4.0 scenario,”
feature extraction and/or dimensionality reduction in the Advances in Manufacturing, vol. 5, pp. 377–387, 2017.
samples collected from the equipment and study the
[12] B. Schmidt and L. Wang, “Predictive maintenance of machine tool linear
computational cost of employing these techniques in 32 bits axes: A case from manufacturing industry,” Procedia manufacturing,
devices. vol. 17, pp. 118–125, 2018.
[13] “Documentação do ASP.NET - Desenvolver aplicativos ASP.NET core,”
From the firmware improvements perspective, the https://learn.microsoft.com/pt-br/aspnet/core/, 2023, accessed:
implementation of an OTA module would be of great value as 2023-06-18.
it would optimize the time spent on firmware update in the [14] “Postgresql: The world’s most advanced open source relational
field. As for the hardware module, the adoption of borne kre database,” https://www.postgresql.org/, 2023, accessed: 2023-06-18.
connector terminals and a hardware reset circuit (watchdog) [15] “O que é o qta?” https://www.ageradora.com.br/o-que-e-qta/, 2023,
accessed: 2023-06-11.
would increase the reliability and robustness of the developed
[16] “Transferência de energia elétrica,” https://apgenergia.com.br, 2023,
prototypes. Another relevant study would be a coverage accessed: 2023-06-11.
analysis of the Mesh network adopted in the project, as well as
[17] "Metrics to evaluate your machine learning algorithm,”
analysis of metrics such as Throughput, Latency and Packet https://towardsdatascience.com, 2023, accessed: 2023-06-18.
Delivery Rate. Thus enabling optimization in the choice of
end nodes that will play the role of root node and child node in
mesh topology.
ACKNOWLEDGMENT
The experiments and results presented in this publication
were obtained through the research and development activities
of the Utility Automation project, a partnership between INDT
- Instituto de Desenvolvimento Tecnológico and ADATA
Eletronics BRAZIL S/A, supported by SUFRAMA under the
terms of Federal Law No. 8248/91.
REFERENCES
[1] J. Dalzochio, R. Kunst, E. Pignaton, A. Binotto, S. Sanyal, J. Favilla,
and J. Barbosa, “Machine learning and reasoning for predictive
maintenance in industry 4.0: Current status and challenges,” Computers
in Industry, vol. 123, p. 103298, 2020.
[2] M. Compare, P. Baraldi, and E. Zio, “Challenges to iot-enabled
predictive maintenance for industry 4.0,” IEEE Internet of Things
Journal, vol. 7, no. 5, pp. 4585–4597, 2019.
[3] T. P. Carvalho, F. A. Soares, R. Vita, R. d. P. Francisco, J. P. Basto, and
S. G. Alcal ́a, “A systematic literature review of machine learning
methods applied to predictive maintenance,” Computers & Industrial
Engineering, vol. 137, p. 106024, 2019.
[4] H. Hua, Y. Li, T. Wang, N. Dong, W. Li, and J. Cao, “Edge computing
with artificial intelligence: A machine learning perspective,” ACM
Computing Surveys, vol. 55, no. 9, pp. 1–35, 2023.
[5] Z. Liu, C. Jin, W. Jin, J. Lee, Z. Zhang, C. Peng, and G. Xu, “Industrial
ai enabled prognostics for high-speed railway systems,” in 2018 IEEE
international conference on prognostics and health management
(ICPHM). IEEE, 2018, pp. 1–8

ISSN 2447-8326. V.1 © 2023 BTSYM