Edge Deep Learning Applied to Granulometric Analysis on

Quasi-particles from the Hybrid Pelletized Sinter (HPS) Process

Nat

alia F. de C. Meira

1 a

, Mateus C. Silva

2 b

, Ricardo A. R. Oliveira

2 c

, Aline Souza

Thiago D’Angelo

and Cl

audio B. Vieira

Metallurgical Engineering Department, Federal University of Ouro Preto, Ouro Preto, Brazil

Department of Computer Science, Federal University of Ouro Preto, Ouro Preto, Brazil

ArcelorMittal, Jo

ao Monlevade, Brazil

Keywords:

AIoT, Artiﬁcial Intelligence, Edge Computing, Edge Learning, Computer Vision.

Abstract:

The mining and metallurgical industry seeks to adapt to Industry 4.0 with the implementation of Artiﬁcial

Intelligence in the processes. The purpose of this paper is to develop the ﬁrst steps of an Artiﬁcial Intelligence

in Deep Learning with Edge Computing to recognize the quasi-particles from the Hybrid Pelletized Sinter

(HPS) process and provide its particle size distribution. We trained our model with the aXeleRate tool using

the Keras-Tensorﬂow framework and the MobileNet architecture and tested it with an embedded system using

the SiPEED MaiX Dock board. Our model obtained 98.60% accuracy in training validation using real and

synthetic images and 100% accuracy in tests with synthetic images and 70% recall. The tests’ results indicate

the feasibility of the proposed system, but with probable overﬁtting in the training stage.

1 INTRODUCTION

The mining and metallurgical sectors are some of the

most traditional productive areas (Kinnunen and Kak-

sonen, 2019). In the later years, innovation and tech-

nology have developed new production and devel-

opment methods (Robben and Wotruba, 2019; Mar-

donova and Choi, 2018; Sinoviev et al., 2016; Chen

et al., 2016; Shibuta et al., 2018). Thus, innovative

projects are crucial for these processes, as they have

a high economic interest. In the steel industry, one of

the primary process parameters is the granulometric

distribution of materials (Zobnin et al., 2018). This

concept means the size distribution of the present par-

ticles, which allows their employability in the produc-

tive process.

When transiting through the production plant, en-

gineers and operators need to know the granulometric

distribution continually. This information is essential

as a process parameter or for making decisions under

critical conditions. Along the steel industry process,

the materials are transported using conveyor belts in

https://orcid.org/0000-0002-7331-6263

https://orcid.org/0000-0003-3717-1906

https://orcid.org/0000-0001-5167-1523

many stages. By itself, the process provokes many

variations in the materials’ features due to humidity,

source of the material, among other factors. These

granulometric distribution changes can jeopardize the

process if they are not within the required speciﬁca-

tions (Januzzi, 2008).

Thus, implementing an algorithm in an embed-

ded system that classiﬁes the grains according to their

granulometric distribution provides a path to solve

this problem and improve the production process. For

this matter, we propose using a deep learning algo-

rithm embedded in an optimized computing device to

classify image samples according to their granulomet-

ric distribution. The usage of artiﬁcial intelligence on

edge computing devices is still an open problem, and

the usage of specialized Edge AI devices allows the

expansion of deep learning towards the IoT (Internet

of Things) (Deng et al., 2020).

In this ﬁrst conjecture, the user must photograph

a sample of the material in the conveyor belt. The re-

sult is accessible through display and also through a

wireless network connection. The implementation of

an Edge Computing solution avoids a large data trans-

mission throughput. This trend pushes the computing

and communication resources to the edge, with faster

Meira, N., Silva, M., Oliveira, R., Souza, A., D’Angelo, T. and Vieira, C.

Edge Deep Learning Applied to Granulometric Analysis on Quasi-particles from the Hybrid Pelletized Sinter (HPS) Process.

DOI: 10.5220/0010458805270535

In Proceedings of the 23rd International Conference on Enterprise Information Systems (ICEIS 2021) - Volume 1, pages 527-535

ISBN: 978-989-758-509-8

527

services and answers to the ﬁnal user (Deng et al.,

2020).

The fast response to detected conditions enables a

better process control. For instance, a granulometry

pattern above the expected is an indicator of elevated

moisture, which can cause clogging in the material

transfer chutes between the conveyor belts. This event

can paralyze the whole production process, exposing

the operators to risk conditions and losing productiv-

ity.

In the industry’s routine, this process can take a

signiﬁcant amount of time and a lack of quality guar-

antee. In many cases, this process takes substantial

time shifts, making it impossible to enable quick re-

sponses due to production variable changes. In the

current applications, the veriﬁcation of certain ma-

terials’ granulometric distribution happens through a

manual process. Figure 1 displays a real sample of

the material obtained from the productive process.

Figure 1: Real Sample example taken from one of the stages

of the steel industry process.

In this task, one operator sample material from the

productive process and manually analyze it in the lab-

oratory to obtain the granulometric distribution. This

procedure happens several times during the day, and

the obtained information is used as a parameter for

the decision making process.

Thus, the manual analysis motivated the develop-

ment of a Deep-Learning-based appliance to detect

quasi-particles. Quasi-particles are material agglom-

erates formed in the HPS process (Januzzi, 2008). We

also embedded this algorithm on a specialized edge

computing device to detect quasi-particles from the

Hybrid Pelletized S

ınter (HPS) process from the steel

industry.

This work is relevant because it consists of imple-

menting a deep learning method in an edge device for

application aimed at the industrial environment, in-

cluding practical tests on embedded hardware.

This paper is organized as follows: In Section 2,

we review the literature and some ground concepts

of this topic. Section 3 presents some of state-of-

the-art the related work. In Section 4, we present a

description of the appliance features, including the

Deep Learning algorithm and the specialized hard-

ware. In Section 5, we explain the employed exper-

imental methodology. The results are presented in

Section 6, and we present further discussions in Sec-

tion 7.

2 THEORETICAL REFERENCES

In this section, we present some theoretical references

about the concepts applied to develop the proposed

solution. This proposal’s main element is a Convo-

lutional Neural Network (CNN) applied to an Edge

Computing solution.

Some of the problems faced in this matter are sim-

ilar to others presented in the literature. For instance,

we observed similar features from this work in pre-

cision agriculture appliances (Keresztes et al., 2018;

Saleem et al., 2019), and even in counting people in

agglomeration (Zhang et al., 2020). Among the pre-

sented challenges, we enforce some aspects:

• Occlusion: Often, quasi-particles overlap one an-

other, causing partial occlusion;

• Complex Background: Homogeneity in shape,

texture, or color from the background and the ob-

jects;

• Rotation: Images are ofter rotated in different an-

gles;

• Illumination Changes: Images are exposed to dif-

ferent levels of light throughout the daytime.

2.1 Deep Learning in Dense Scenes

Lecun et al. (LeCun et al., 2015) state that Deep

Learning (DL) is a set of techniques from the Ma-

chine Learning universe, often referred to as Arti-

ﬁcial Intelligence. These algorithms’ formalization

comes from the Artiﬁcial Neural Networks (ANN),

containing multiple hidden layers and massive train-

ing datasets. According to Zhang et al. (Zhang et al.,

2020), DL algorithms represent state of the art on Ma-

chine Learning techniques. Nonetheless, the detec-

tion of objects in dense scenes is particularly chal-

lenging.

Zhang et al. (Zhang et al., 2020) separate dense

scenes into two different classes: quantity dense

scenes and internally dense scenes. In the ﬁrst one,

there is a large number of objects of interest in the

scene. The second one happens when the objects have

dense inner attributes. In both cases, labeling the data

ICEIS 2021 - 23rd International Conference on Enterprise Information Systems

528

is a signiﬁcant challenge, as the classiﬁcation is af-

fected by noise and resolution on small objects de-

tection. According to these authors, the best DL ar-

chitectures for classiﬁcation in dense scenes are VG-

GNet, GoogLeNet, ResNet. Also, the best architec-

ture for object detection are DetectNet and YOLO.

Gao et al. (Gao et al., 2020) analyzed 220 re-

lated works to understand the crowd counting pro-

cess systematically. These authors point out that the

main challenge is the detection of small objects in a

scene. This trait happens as in crowd scenes, the in-

dividuals’ heads are often too small. According to the

authors, the most successful techniques for counting

crowds based on detection are SSD, YOLO, and R-

CNNs. Although these architectures had success in

sparse scenes, these networks had unsatisfactory re-

sults given scenes with occlusion, disorder, and dense

background. Furthermore, SSD is not efﬁcient with

small objects on the images, as its intermediate layers

resource mapping may dilute the detected object’s in-

formation. For the R-CNN, Zhou et al. (Zhou et al.,

2019a) proposed an improvement based on PCA Jit-

tering to enhance the detection of small objects on the

Faster R-CNN architecture.

The presented work display some of the chal-

lenges in developing Convolutional Neural Networks

(CNNs) capable of analyzing dense scenes with oc-

cluded objects. This issue is more signiﬁcant when

the dataset complexity increases. Developers often

follow a synthetic database procedure to solve this

problem, with further validation with actual real data.

The obtained results are usually good, except if there

is a substantial deviation from the synthetic and real

datasets (Zhang et al., 2020).

2.2 Edge AI

Another critical aspect of the solution is the algorithm

persistence in edge computing applications. The evo-

lution of embedded computing technologies raises the

challenge of providing machine learning as services

in edge applications with quality. Thus, the creation

of reduced models and specialized hardware create

the concept of an “Edge AI” (Wang et al., 2019). This

novel perspective targets using machine learning in

edge devices with independence from cloud applica-

tions.

Nonetheless, developing machine learning and es-

pecially DL models for edge computing devices is

a challenging task. Deep Neural Networks (DNNs)

usually are computationally intensive models (Li

et al., 2019). On the one hand, DNNs usually require

much computational power. On the other hand, mov-

ing this application to the cloud requires a high data

throughput through a network infrastructure. The in-

creasing number of devices can easily exceed the net-

working capacities (Lin et al., 2019).

Zhou et al. (Zhou et al., 2019b) state that there are

some issues to solve for enabling the Edge AI devel-

opment. Among these challenges, we enforce:

• Programming and Software Platforms;

• Resource-Friendly Edge AI;

• Computational-Aware Techniques;

Another aspect to be considered in the development

of novel edge computing solutions is the hardware

constraints. As stated before, most DL architectures

require a high computational charge. An outcome

for this problem is integrating dedicated hardware

to optimize Edge AI solutions (Mazzia et al., 2020;

Ohbuchi, 2018; Karras et al., 2020).

3 RELATED WORK

Given the importance of the iron ore agglomeration

stage for the later stages of the process, several stud-

ies have been carried out to control and monitor the

variables that interfere in the sintering and pelletizing

processes.

Dias (Dias, 2018) proposed a granulometric con-

trol system for iron ore pellets by controlling the wa-

ter injection in the pellet drum, which, until then, was

done manually by the operators according to the need

of the process. The results showed that water addition

tends to increase the pellets’ granulometry and that

the control tends to homogenize the pellets. However,

for the controlled variable to present stabilization, it

would be necessary to study other parameters, such

as water saturation due to pellet recirculation outside

the required particle size range.

Studies on the inﬂuence of raw materials in the

cold agglomeration process of the HPS process were

also studied, as shown in Januzzi (Januzzi, 2008). The

work had the objective to characterize the raw mate-

rials, study the contribution of each of them in the

cold agglomeration process, and adjust the parame-

ters to improve the process’s performance. One of the

measures taken was the changes in the granulomet-

ric distribution curves of serpentinite, limestone, and

manganese ore, which promoted an improvement in

the quasi-particles’ average size. Consequently, this

measure causes “a positive effect on the suction pres-

sure in the sinter allowing the increase of layer height,

gain in productivity and sinter production” (Januzzi,

2008), once again demonstrating the importance of

granulometric distribution in the iron ore agglomera-

tion process.

Edge Deep Learning Applied to Granulometric Analysis on Quasi-particles from the Hybrid Pelletized Sinter (HPS) Process

529

For the case where the manual control depended

on the area operators to obtain the adequate granu-

lometry of the raw pellet, Passos et al. (Passos et al.,

2014) developed its work in the implementation of an

advanced control system (SCAP) intending to control

the granulometry of the pellets raw materials acting

on the speed and feeding of the disks. The results

showed the stability of the production process, mainly

in controlling the pellets’ granulometric distribution,

the stability of the dosage of inputs, and the hardening

furnace’s increased permeability.

To characterize ultraﬁne materials and average

size consumption, Gontijo (Gontijo, 2018) performed

previous image treatment using the Scanning Elec-

tronic Microscope (SEM). The particles in the im-

ages were digitized, scaled using software, classiﬁed

by colors in size ranges (intervals), and, after clas-

siﬁcation, graphs of granulometric distributions were

generated.

4 EDGE AI HARDWARE

PRESENTATION

In this work, we decided to implement the solution us-

ing the SiPEED MAiX Dock board, displayed in Fig-

ure 2. Some performance numbers of the board are

shown in Table 1. The work of Klippel et al. (Klip-

pel et al., 2020) demonstrates the comparison between

SiPEED MaiX BiT, Raspberry Pi 3, and Jetson Nvidia

Nano cards. The authors implemented the SiPEED

MaiX BiT for the detection of tears in conveyor belts.

The SiPEED MaiX Dock board is similar to the one

used in this work, and we follow the methodology

proposed by Klippel et al.

This platform has an onboard device with artiﬁcial

intelligence (AI) hardware acceleration. MAiX is the

module explicitly developed for SiPEED, designed to

perform AI. It offers high performance considering a

small physical and energy area, allowing the implan-

tation of high precision AI and a competitive price.

The main advantages of this device are:

• Complete hardware and software infrastructure to

facilitate the deployment of AI-based solutions;

• Good performance, small size, low energy con-

sumption, and low cost, which allows a broad de-

ployment of high quality AI on board;

• It can be used for an increasing number of indus-

trial use cases, such as predictive maintenance,

anomaly detection, machine vision, robotics, and

voice recognition.

The SiPEED MAiX acts as the master controller,

and the hardware has a KPU K210. MaixPy is a

Figure 2: SiPEED M1 Dock.

Table 1: Embedded platform performance numbers.

Parameter Characteristics

CPU 64-bit RISC-v processor and core

Chipset K210 - RISC - V

Image Recognition qvg at 60fps / vg at 30fps

Clock (GHz) 0.40

AI resources KPU

OS/Language uPython

Dimensions (mm) 60x43x5

framework designed for AIoT programming, prepare

on an AIoT K210 chip, and based on the Micropy-

thon syntax. MicroPython is a lean and efﬁcient

implementation of the Python 3 programming lan-

guage, which includes a small subset of the standard

Python library, and is optimized to run on micro-

controllers and in restricted environments, facilitating

programming on the K210 hardware. MAiX supports

a ﬁxed-point model that a conventional training struc-

ture trains according to speciﬁc restriction rules and

has a model compiler to compile models in its model

format. It is compatible with network architectures

Tiny-Yolo and MobileNet-v1.

The Kendryte K210 is a dual-core RISCV64 SoC

with AI capability that has machine vision capabili-

ties and can perform low energy consumption Convo-

lutional Neural Networks (CNNs) calculations, with

features for object detection, image classiﬁcation, de-

tection and face recognition, obtaining target size and

coordinates in real-time and obtaining the type of tar-

get detected in real-time. Figure 3 displays the K210

block diagram (ken, 2018). The KPU is a general-

purpose neural network processor with internal con-

volution, normalization, activation, and pooling oper-

ations. According to the manufacturer, it also has the

following characteristics:

• Supports the ﬁxed-point model that the conven-

tional training structure trains according to spe-

ciﬁc restriction rules;

• There is no direct limit on the number of network

layers, and each layer of the convolutional neural

network parameters can be conﬁgured separately,

including the number of input and output chan-

nels, the width of the input and output line, and

the height of the column;

• Support for 1x1 and 3x3 convolution kernels;

ICEIS 2021 - 23rd International Conference on Enterprise Information Systems

530

• Support for any form of activation function;

• The maximum size of the supported neural net-

work parameter for real-time work is from 5MiB

to 5.9MiB.

Figure 3: Block diagram of K210.

For training, the aXeleRaTe framework, a Keras-

based framework for AI on the Edge, was used to run

computer vision applications (image classiﬁcation,

object detection, semantic segmentation) on edge de-

vices with hardware acceleration. AXeleRate simpli-

ﬁes the training and conversion of computer vision

models and is optimized for workﬂow on the local

machine and Google Colab. Supports conversion of

trained model to: .kmodel (K210) and .tﬂite formats.

Figure 4 displays the process of using aXeleRate,

with the main steps indicated by the blue circles. In

(1), the dataset is loaded from Google Drive for train-

ing in the Keras-Tensorﬂow framework. Then (2), the

model is delivered in the .h5 format for classiﬁcation

and returns to Tensorﬂow (3) to be converted into the

.tﬂite format (4). Thus, it is delivered to nncase (5)

to be compiled into the format .kmodel (6), which is

executed by KPU (7).

This work’s main contribution is the implementa-

tion of a deep learning method on an edge device for

application aimed at the industrial environment, in-

cluding practical tests on embedded hardware.

5 EXPERIMENTAL

METHODOLOGY

This section assesses the experimental methodology

used to validate the appliance, given the targeted

hardware. For this matter, we present the employed

Figure 4: Training and compilation with aXeleRate.

dataset, training process, and evaluation metrics. We

test a pilot application classiﬁer’s performance and

validate the model’s transfer into the desired hard-

ware.

5.1 Dataset

We generated a dataset with 1368 images to create a

pilot appliance, containing 1140 for training and 228

for validation. The dataset has three different classes:

quasi-particle, non-category, and empty. We also

added 343 synthetic images produced on the bench-

scale for the quasi-particle class training, as presented

in Figure 5. These images were generated to avoid

the problems of overlapping and occlusion of the par-

ticles. We also added another 343 images of sam-

ples of quasi-particles carried out in a company in

the mining-metallurgical sector with real data to con-

tribute to the quasi-particle training dataset.

Figure 5: Image of dataset developed on bench-scale.

5.2 Training the Deep Learning Model

We conducted the training of the deep learning model

on the Google Collaboratory platform. This process

was carried out using the aXeleRate

tool. This appli-

cation is a tool for training classiﬁcation and detection

https://github.com/AIWintermuteAI/aXeleRate

Edge Deep Learning Applied to Granulometric Analysis on Quasi-particles from the Hybrid Pelletized Sinter (HPS) Process

531

models developed using the Keras/Tensorﬂow frame-

work.

To perform the desired task, we chose to use the

MobileNet as CNN architecture. We used version

0.75 MobileNet-224 v1, conﬁgured as a classiﬁer,

with 224 inputs, two layers fully connected with 100

and 50 neurons, and a dropout of 0.5. The training

session held thirty training seasons, and the learning

rate adopted was 0.001. The initial weights of the

model were loaded, considering the previous training

with the ImageNet dataset. Also, data augmentation

was performed during the training.

5.3 Evaluation Metrics

At ﬁrst, the classiﬁcation model’s performance was

calculated using the Confusion Matrix, which shows

the classiﬁcation frequencies for each class of the

model. From this data, we extract the parameters:

precision, given by (1), recall given by (2) and overall

accuracy F1, given by (3). These parameters deﬁne

how well the model worked, how good the model is

for predicting positives, and the balance between the

precision and the recall of the model.

For this matter, we followed the presented deﬁ-

nitions: TP is a true-positive sample, FP is a false-

positive sample, TN is a true-negative sample, and

FN is a false-negative. TP occurs when the main class

prediction is correct, and FP when it is mispredicted.

TN occurs when the alternative class prediction is cor-

rect and FN when it is mispredicted.

precision =

T P

T P + FP

(1)

recall =

T P

T P + FN

(2)

F1 = 2 ∗

precision * recall

precision + recall

(3)

5.4 Edge AI Construction

We assembled a SiPEED Dock plate for the execution

of the bench-scale model with synthetic images. For

this test, we used two Python scripts used for the tests.

The ﬁrst to capture photos with 224x224 resolution

and storage on the SD card. The second to test the

model from the storage data set previously stored on

the SD card.

6 RESULTS

We present here the obtained results from the appli-

cation of this procedure. Our preliminary results indi-

cate the system feasibility and show the constraints to

transport the model into the Edge AI device.

6.1 Training Model Performance

The training elapsed time was 54 minutes, reaching

an accuracy of 98.60%. Figure 6 displays the evolu-

tion of the accuracy throughout the training stage. As

displayed in the graph, the model’s training converged

in just ten iterations, indicating that the model had no

great difﬁculty in differentiating the classes of images

present in the database.

Figure 6: Training graph.

To validate the model, we created a dataset with 228

images. These frames were divided into three classes,

containing 76 images each class: quasi-particle, non-

category, and empty (empty refers to the same tray,

but without the presence of quasi-particles). Table

2 displays the confusion matrix considering quasi-

particles as the main class and Table 3 shows the per-

formance indicators.

Table 2: Confusion matrix of model - Validation set.

Predict

quasi particle non category empty

quasi particle 76 0 0

Real non category 1 74 1

empty 0 0 76

Table 3: Trained Model Performance at Validation set.

Indicator Value

precision 98,60%

recall 100%

F1 99,34%

The model accuracy was 98.70%. The application

displayed problems in classifying some uncategorized

ICEIS 2021 - 23rd International Conference on Enterprise Information Systems

532

images with quasi-particle and empty trays. The data

suggest a good recall, which means that the model had

a small error rate in the quasi-particles’ classiﬁcation

when they were indeed quasi-particles. These results

demonstrate the feasibility of the recognition process

using the proposed dataset. This value enabled a bal-

ance in the F1 score.

6.2 Model Performance at Edge AI

We also tested the performance of the classiﬁer in the

edge computing candidate platform. After training,

we loaded the model into the SiPEED Maix Dock for

testing, as showed in Figure 7. For this matter, we

tested the system using images from the three classes

(quasi-particle, non-category, and empty). Table 4

displays the confusion matrix and Table 5 shows the

performance indicators.

Figure 7: SiPEED MaiX Dock - Test Demonstration.

Table 4: Confusion matrix of model - Test set.

Predict

quasi particle non category empty

quasi particle 7 2 1

Real non category 0 9 1

empty 0 0 10

Table 5: Trained Model Performance at Test set.

Indicator Value

precision 100,00%

recall 70,00%

F1 82,35%

In contrast to the value achieved in the validation set,

the recall in the test set dropped to 70%, assessed from

the SiPEED embedded system. This result indicates

that the model had difﬁculties in testing positive for

simulations of images similar to images in the indus-

trial environment, as displayed in Figures 9 and 8, al-

though for synthetic images with spaced particles was

no difﬁculty, as shown in Figure 10.

The work of Klippel et al. (Klippel et al., 2020)

implemented the SiPEED MaiX BiT to detect fail-

ures in conveyor belts. Our results for training perfor-

mance are similar to the results obtained by Klippel

et al. In the test performance, we obtained a lower

recall, as shown.

The recall value in the tests does not match the

results obtained in the tests carried out by Klippel et

al.(Klippel et al., 2020) To justify the value of 70 %,

we understand that the data set can be improved to

only real images in future analyses. Also, there is

a possibility of overﬁtting during training. In order

to verify this hypothesis, we intend to increase the

database in future works.

These data demonstrate the difﬁculty of reconcil-

ing results obtained on a bench scale with results close

to real environments.

Figure 8: Example of recognition of quasi-particles simu-

lating sampling in an industrial environment during the test

using SiPEED.

Figure 9: Example of error in recognizing quasi-particles

simulating sampling in an industrial environment during the

test using SiPEED.

Edge Deep Learning Applied to Granulometric Analysis on Quasi-particles from the Hybrid Pelletized Sinter (HPS) Process

533

Figure 10: Example of recognition of quasi-particles with

sample developed on a bench scale during the test using

SiPEED.

7 CONCLUSION

In this work, we proposed an Edge AI appliance to

classify images from the Hybrid Pellet Sinter (HPS)

process from the steel industry. For this matter, we

produced an application that aimed to detect trays

containing quasi-particles, allowing them to differen-

tiate from other objects and even empty trays. The

results indicate the system feasibility and the possi-

bility of loading the produced model into an Edge AI

specialized hardware, although it needs improvement.

The importance of obtaining the intervals of the par-

ticle size distribution of quasi-particles in the HPS

process, its use of edge is shown as an advance, as

currently, this process is carried out manually and in

prolonged time intervals in the plants.

We proposed the appliance of a Convolutional

Neural Network (CNN) embedded into an Edge Com-

puting appliance. For this matter, the algorithm must

analyze a dense scene, with problems like occlusion

and complex background changes. Although there are

broad applications of deep learning in dense scenes,

there are still open issues to solve in the research

process. The usage of specialized edge computing

hardware enhances the performance of the solution on

edge.

For testing this application, we used the SiPEED

MaiX Dock board. This board has hardware and

software infrastructure to enhance the development

of Edge AI applications. The solution is cost- and

resource-restrictive, given its specialized hardware to

create edge deep learning applications. It also inte-

grates with models created using the main DL frame-

works.

Tests using SiPEED allow the detection of quasi-

particles in synthetic images without difﬁculties, even

with particles and their spaced distribution. However,

tests with real images obtained some ﬂaws, evidence

by the drop in recall to 70%. There may have been

overﬁtting during or training or, during the tests, had

inﬂuences associated mainly with the brightness of

the day, the occlusion between particles, color homo-

geneity, and overlap between objects. These problems

have been reported in other works with previous ad-

vanced problems.

From the results obtained in this step, it was possi-

ble to raise new hypotheses for approaches to improve

the deep learning algorithm to provide a granulomet-

ric range of the quasi-particles present in a sample.

Thus, it is intended to evaluate other models for future

work, such as supplying instances, making it possible

to extract the histogram of the particle size distribu-

tion from the images containing quasi-particles. The

development of an edge device with deep learning can

bring signiﬁcant beneﬁts, both from process improve-

ment and the insertion of steel processes in Industry

4.0.

ACKNOWLEDGEMENTS

The authors would like to thank CAPES, CNPq and

the Federal University of Ouro Preto for supporting

this work. Also, the authors would like to thank

ArcelorMittal Monlevade for enabling the creation of

a dataset with real images. This study was ﬁnanced in

part by the Coordenac¸

ao de Aperfeic¸oamento de Pes-

soal de N

ıvel Superior - Brasil (CAPES) - Finance

Code 001.

REFERENCES

(2018). K210 datasheet.

Chen, Z., Liu, L., Qi, X., and Geng, J. (2016). Digital min-

ing technology-based teaching mode for mining engi-

neering. iJET, 11(10):47–52.

Deng, S., Zhao, H., Fang, W., Yin, J., Dustdar, S., and

Zomaya, A. Y. (2020). Edge intelligence: the con-

ﬂuence of edge computing and artiﬁcial intelligence.

IEEE Internet of Things Journal.

Deng, S., Zhao, H., Fang, W., Yin, J., Dustdar, S., and

Zomaya, A. Y. (2020). Edge intelligence: The con-

ﬂuence of edge computing and artiﬁcial intelligence.

IEEE Internet of Things Journal, pages 1–1.

Dias,

I. d. S. M. (2018). Sistema de controle granulom

etrico

de pelotas de min

erio de ferro.

Gao, G., Gao, J., Liu, Q., Wang, Q., and Wang, Y. (2020).

Cnn-based density estimation and crowd counting: A

survey. arXiv preprint arXiv:2003.12783.

Gontijo, M. D. (2018). An

alise granulom

etrica por imagem

ICEIS 2021 - 23rd International Conference on Enterprise Information Systems

534

de amostras ultraﬁnas. Revista Engenharia de Inter-

esse Social, 1(3).

Januzzi, A. (2008). An

alise da aglomerac¸

ao a frio no pro-

cesso hps (hybrid pelletized sinter) com

enfase nas

mat

erias-primas envolvidas.

Karras, K., Pallis, E., Mastorakis, G., Nikoloudakis, Y.,

Batalla, J. M., Mavromoustakis, C. X., and Markakis,

E. (2020). A hardware acceleration platform for ai-

based inference at the edge. Circuits, Systems, and

Signal Processing, 39(2):1059–1070.

Keresztes, B., Abdelghafour, F., Randriamanga, D.,

da Costa, J.-P., and Germain, C. (2018). Real-time

fruit detection using deep neural networks. In 14th

International Conference on Precision Agriculture.

Kinnunen, P. H.-M. and Kaksonen, A. H. (2019). Towards

circular economy in mining: Opportunities and bot-

tlenecks for tailings valorization. Journal of Cleaner

Production, 228:153–160.

Klippel, E., Oliveira, R., Maslov, D., Bianchi, A., Silva,

S. E., and Garrocho, C. (2020). Towards to an embed-

ded edge ai implementation for longitudinal rip de-

tection in conveyor belt. In Anais Estendidos do X

Simp

osio Brasileiro de Engenharia de Sistemas Com-

putacionais, pages 97–102, Porto Alegre, RS, Brasil.

SBC.

LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learn-

ing. nature, 521(7553):436–444.

Li, E., Zeng, L., Zhou, Z., and Chen, X. (2019). Edge

ai: On-demand accelerating deep neural network in-

ference via edge computing. IEEE Transactions on

Wireless Communications, 19(1):447–457.

Lin, X., Li, J., Wu, J., Liang, H., and Yang, W. (2019). Mak-

ing knowledge tradable in edge-ai enabled iot: A con-

sortium blockchain-based efﬁcient and incentive ap-

proach. IEEE Transactions on Industrial Informatics,

15(12):6367–6378.

Mardonova, M. and Choi, Y. (2018). Review of wearable

device technology and its applications to the mining

industry. Energies, 11(3):547.

Mazzia, V., Khaliq, A., Salvetti, F., and Chiaberge, M.

(2020). Real-time apple detection system using em-

bedded systems with hardware accelerators: An edge

ai application. IEEE Access, 8:9102–9114.

Ohbuchi, E. (2018). Low power ai hardware platform

for deep learning in edge computing. In 2018 IEEE

CPMT Symposium Japan (ICSJ), pages 89–90. IEEE.

Passos, L. A. S., Moreira, J. L., Jorge, A., and Cavalcante,

M. V. S. (2014). Melhoria no desempenho do processo

de produc¸

ao de pelotas de min

erio de ferro em discos

de pelotizac¸

ao pela utilizac¸

ao de sistemas otimizantes

com l

ogica nebulosa. In ABM Proceedings. Editora

Blucher.

Robben, C. and Wotruba, H. (2019). Sensor-based ore sort-

ing technology in mining—past, present and future.

Minerals, 9(9):523.

Saleem, M. H., Potgieter, J., and Arif, K. M. (2019). Plant

disease detection and classiﬁcation by deep learning.

Plants, 8(11):468.

Shibuta, Y., Ohno, M., and Takaki, T. (2018). Com-

putational metallurgy: Advent of cross-scale mod-

eling: High-performance computing of solidiﬁcation

and grain growth (adv. theory simul. 9/2018). Ad-

vanced Theory and Simulations, 1(9):1870020.

Sinoviev, V. V., Okolnishnikov, V. V., Starodubov, A. N.,

and Dorofeev, M. U. (2016). Approach to effective-

ness evaluation of robotics technology in mining us-

ing discrete event simulation. International Journal of

Mathematics and Computers in Simulation, 10:123–

128.

Wang, X., Han, Y., Wang, C., Zhao, Q., Chen, X., and Chen,

M. (2019). In-edge ai: Intelligentizing mobile edge

computing, caching and communication by federated

learning. IEEE Network, 33(5):156–165.

Zhang, Q., Liu, Y., Gong, C., Chen, Y., and Yu, H. (2020).

Applications of deep learning for dense scenes analy-

sis in agriculture: A review. Sensors, 20(5):1520.

Zhou, X., Fang, B., Qian, J., Xie, G., Deng, B., and Qian,

J. (2019a). Data driven faster r-cnn for transmission

line object detection. In Cyberspace Data and In-

telligence, and Cyber-Living, Syndrome, and Health,

pages 379–389. Springer.

Zhou, Z., Chen, X., Li, E., Zeng, L., Luo, K., and Zhang, J.

(2019b). Edge intelligence: Paving the last mile of ar-

tiﬁcial intelligence with edge computing. Proceedings

of the IEEE, 107(8):1738–1762.

Zobnin, N. N., Torgovets, A. K., Pikalova, I. A., Yussupova,

Y. S., and Atakishiyev, S. A. (2018). Inﬂuence of ther-

mal stability of quartz and the particle size distribution

of burden materials on the process of electrothermal

smelting of metallurgical silicon. Oriental Journal of

Chemistry, 34(2):1120–1125.

Edge Deep Learning Applied to Granulometric Analysis on Quasi-particles from the Hybrid Pelletized Sinter (HPS) Process

535