In this project, we present the design of a low-noise amplifier in complementary metaloxide-semiconductor (CMOS) 180nm technology. This amplifier is integrated into a passive Ultra High Frequency (UHF)/Ultra Wideband (UWB) tag with the purpose of monitoring cardiac and pulmonary blood flow in a patient’s residence. Considering the significant advancements in integrated circuits over recent years, spanning a wide range of applications in analog and digital electronics, it is asserted that the UWB standard is a suitable choice for non-invasive biomedical applications and various applications in the Internet of Things (IoT) in general. The tag performs measurements, data transmission, and reception of vital information, enabling various forms of control and alarm generation. The circuit is powered through energy harvesting from the UHF band. Additionally, a power management unit is integrated to control battery usage when data reception is active, also powered by energy harvesting. The primary function of the low-noise amplifier (LNA) in the UWB receiver is to increase the signal amplitude with minimal distortion and control the noise figure of the entire receiver. Thus, the LNA project was evaluated based on reflection losses at the input and output, gain, noise figure, stability, and consumed area. Corner simulations were conducted considering extreme variations in temperature and supply voltage, along with post-layout simulations after parasitic extraction.

In Brazil, soybean production has increased considerably in recent decades, driven by ad-
vanced agricultural technologies. However, the excessive use of pesticides, which represents
an important portion of the costs, presents economic and environmental challenges. Pest
detection and counting methods using Computer Vision appear promising, since traditional
approaches are laborious and time-consuming. This study introduces a method tailored for
real-time detecting and counting insects in soybean fields, built upon an improved YOLOv8
model, designed for this mission. To enhance the accuracy in detecting small insects and
reduce the complexity of YOLOv8, we conducted ablation experiments to assess the impact
of integrating a deeper feature level and a proposed C2f2 layer into the insect detection model.
Through these ablation experiments, new algorithm underwent training and testing using a
public research dataset, composed of samples of Euschistus Heros (NBSB), a bug of interest
in soybean crops in Brazil, and its performance was compared against three configurations
under identical conditions: A) YOLOv8n, B) YOLOv8n with C2f2 only and C) YOLOv8n
with P2 only. Our evaluation employed various metrics, including Precision, mAP0.5 and
mAP0.95. We also considered model complexity as essential factors in assessing the effi-
ciency of YOLO models for specific applications, by comparing Params, FLOPs, Inference,
and Time. The proposed model was then integrated into a framework capable of track and
count NBSBs from a video stream. This video stream was created, animating 42 new images,
captured under diverse lighting and background conditions to address potential challenges
in practical applications, testing the generalization capacity of the model as well as the
model’s performance in video applications. The model demonstrates a satisfactory measure
of accuracy, represented by the MOTA metric, with around 62%. This indicates that the
proposed technique is effective in tracking bedbugs well throughout the scenes presented,
which is corroborated by the high hit rate in the final insect count (upward deviation of
5.3%). These results point to the potential of this type of framework and signal potential in
future pest mapping applications based on real-time tracking.

In a competitive global environment, maintenance plays a fundamental role in the availability of equipment in the food industry. This study seeks to balance maintenance strategies to optimize equipment uptime, developing a preventive plan based on quantitative analyses, including MTBF (MeanTime Between Failures) and MTTR (Mean Time To Repair). ) and availability of equipment and system. In the methodological approach adopted in this study, 11 models of industrial air conditioners were analyzed based on performance and efficiency criteria. Historical data from these equipment was used to calculate MTTR, MTBF and availability. A preventive maintenance plan was developed based on these results. The approach followed statistical modeling using Minitab® version 16 software and the phases of the data analysis reference model. The Anderson-Darling test was used to evaluate data adherence to different probabilistic distributions. The sample consisted of 399 observations of repair time and time between failure data. Probability distributions, such as Exponential, Gamma, LogNormal and Weibull, and the approach to evaluating the preventive maintenance strategy were detailed. In this study, we explored the importance of preventive maintenance in the food industry, focusing on the reliability and availability of air conditioning equipment. In this study, historical data and statistical modeling are used to analyze the mean time between failures (MTBF), the mean time to repair (MTTR) and equipment availability. Developed a preventive maintenance plan based on these average times and availability, aiming to optimize production efficiency and effectiveness. Given the analyzes carried out, it is clear that the implementation of preventive maintenance strategies is essential to guarantee the reliability and availability of equipment in the food industry. Statistical modeling of the data made it possible to evaluate the performance of the equipment over time, enabling the creation of an effective maintenance plan. It is recommended that these preventive practices continue to mitigate the impacts of failures in production processes and guarantee the quality and integrity of food products.

This thesis explores essential concepts to justify an innovative research proposal in chaotic secure communication. Initially, a comprehensive review of articles and surveys related to secure communication was conducted, focusing on chaotic systems, actuated control, robustness, and hyperchaotic systems. This provided an understanding of the current state of research in communication security. Next, a statistical analysis was conducted to evaluate recent communication system models, considering simplicity, robustness, security, and reliability in encoding and decoding encrypted messages, as well as the ability to handle disturbances. The methodology involved collecting data from reliable sources such as Google Scholar and Scopus. The results highlight a growth in publications, with significant leadership from China. Additionally, trends in hyperchaotic secure communication will be analyzed. Following this, an innovative secure communication system is proposed, using chaos-based analog cryptography to transmit signals in a public range. The main contributions include the introduction of a subactuated control law in synchronization, demonstrating system robustness against disturbances, and exploring unpredictability in hyperchaotic systems to enhance cryptography security. These contributions consolidate the proposed system as a promising approach to improving security in communication within public environments, providing a solid foundation for future research and practical applications in secure telecommunications.

Infusion Pumps (IP) are medical devices (MD) used to administer medicines or food continuously and precisely. Over the years, its use has expanded, currently being used in emergencies, ICUs, pediatrics and several other hospital sectors. In the recent COVID-19 pandemic, there has been an intensification of the use of this device due to the demand for the administration of high doses of sedatives for ventilated patients. As it is an invasive medical equipment, maintenance is required with greater attention and frequency, in order to maintain good functioning and precision in the infusion. If the IPs fail, they can cause adverse events that can cause harm to the patient, harming their health. In addition to the problems associated with the patient, the IP is the MD that has the highest number of adverse event reports in Brazil, which increases the need to verify the conditions under which the device is operating, aiming to minimize failures during use.

Given this panorama, the objective of this research is to analyze the reliability of IP operating in a Brazilian hospital, using an internal database, built from Clinical Engineering software. Probability distributions for repair time and time between failures were modeled, aspects such as reliability and availability of hospital IPs were calculated through failure analysis, in accordance with national/international recommendations, and hospital sectors with recurrent failures were investigated. and services performed. Furthermore, we propose to carry out a systematic literature review to certify how studies on the reliability of pumps and other metrological parameters have been studied in recent years, to finally propose a plan for checking the pumps.

The review demonstrated that the gravimetric method is a proposal that brings positive results in flow assessment, being an alternative for those responsible for clinical engineering. It was evident that there is a lack of studies that address reliability concepts applied to IP, highlighting those involving bench and on-site tests. In the evaluation of the operating equipment, it was observed that there were no details in the notes regarding the reason for the failure and when opening the work order, which could lead to inaccuracy in maintenance planning. The longest repair time was identified in the Intensive Care Unit (Neurological), a sector that comprises the majority of pumps. Graphical analyzes and tests showed that the Weibull distribution satisfactorily models both the time between failures and the time to repair machines, and model A demonstrated better results in terms of availability and reliability. Finally, the study proposes a plan that combines national recommendations with a reliability cycle for DM aimed at verifying BI, so that it can be included in the predictive and preventive maintenance routine.

The main objective of this work is the development of a protocol and the modification of the ARFACTA radiofrequency ablation equipment to obtain the values of the components of the cole-cole model, which will be used in the Hardware in the loop (HIL) module and in the controller. From this modification, a test bench will be created that will use the HIL technology to emulate the behaviour of an organ during the ablation procedure.

This test bench will be able to represent the impedance variation of the organ that has undergone the treatment, allowing to verify the behaviour and evaluate the effectiveness of the process. The device will enable controller testing, seeking to maximise treatment time by delaying the roll-off of the frequency response.

The development includes the creation of a frequency controller with the purpose of optimising the ablation parameters, enabling the modification of the output frequency in a pre-established range. This dynamic controller has the ability to cover different treatment regions, covering the variables inherent to the organ in question and to the treatment process itself. This advancement will allow maximising the treatment time by delaying the roll-off of the frequency response, aiming to improve the effectiveness of the medical procedure and achieve more satisfactory outcomes for patients.

This set of developments will contribute to the advancement of ablation technology, providing greater efficiency and control in organ treatment, as well as enabling more accurate evaluation of the performance indicators of the controllers used. With this, it is expected to improve the effectiveness of medical procedures, ensuring more satisfactory results for patients.

The present work presents the development of a proportional integral derivative (PID) tuning for a motor generator set using bioinspired algorithms, and proposes an alternative to meet the tuning in offshore environments. The tuning of the controller was carried out from the formulation of an optimization problem, whose performance analysis is done by tracking the trajectory of the response in time domain, applying disturbances when increasing and decreasing loads in the power system, thus providing the necessary data to optimize the system and raise the mechanical potential, reducing instabilities and maintenance and increasing the energy efficiency. A preliminary analysis of the components and variables involved in the motor-generator system is presented, namely: internal combustion engine, electric generator, controller and data acquisition system. A simplified non-linear modeling was made, from a transfer function of a servo system. The obtained model was used to extract a set of controller performance parameters that guide the search process of particle swarm optimization (PSO) and differential evolution (DE) algorithms, as well as their variations, using Matlab/Simulink. For comparative purposes, the PIDTune function was used, allowing the evaluation of the performance of bioinspired algorithms. With the obtained PID parameters, project simulations and discussions were carried out. Subsequently, the PID parameters were implemented in the controller, following trajectory tracking tests with a Wartsila engine, coupled to a Leroy Somer electric generator from the Petrobras platform 65 offshore installation. Finally, a discussion is made about the effectiveness of the proposed optimization of the controller with the selected tuning and an analysis of the engine-generator set parameters is presented, as well as conclusions and suggestions for further research.

In Brazil, the information systems aimed at meeting social demands such as public health services have had a more concerned look at the global health crisis - COVID-19- that we are facing, but there are still limitations in the field of information technology in public health services, primary care, verification of scheduled exams, consultations, and care in vaccine rooms. This chapter will deal with the Research and Innovation Project, the 3TS that comprises the SUS+ modules, which is the possibility of self-service to the patient, and the Imuna SUS that shows the control of received vaccines. This set of Technology and Communication Systems (TIC) has steps that go from the original conception of the information system to its arrival at SUS, adding value to the services provided by SUS

Problem: Despite the great advances in modeling in radiofrequency ablation, a great difficulty is due to the intrinsic characteristics of human body tissues. Variations in the amount of water and fat in the body composition of patients generate significant variations between the models and the results obtained in the ablations.

Objective: To develop a generalist radiofrequency ablation equipment, capable of performing complex bioimpedance reading and adjusting the frequency and amplitude of the wave, to improve impedance matching and increase the efficiency of the ablation process.

Methodology: The development was based on the circuit of the equipment that was already under development, but all circuits were analyzed, simulated and some were designed. The equipment has a main source, with a voltage regulator circuit for 5 and 12V dc (Volts in direct current), has a variable Buck converter, which is responsible for feeding the power circuit, a Push-Pull circuit to generate the alternating signal, current and voltage measurement circuits, in addition to the phase shift measurement circuit, so that it can determine the complex impedance of the tissue. To validate the operation of the equipment, we performed ablation in porcine and bovine liver tissue with power variation.

Results: The ablation performed in bovine liver presented a curve relating the variation of power with the ablation area, showing that there is an optimal time for ablation with constant power to be achieved. The tests on porcine liver showed constancy in the ablation area, even though the tests were performed in different sectors. It was also possible to determine that the implementation of a control is able to significantly increase the ablation area.

Conclusion: The equipment was able to receive voltage and frequency control techniques to maximize power transfer and increase the ablation area. Due to the ability of the equipment to communicate via wi-fi, there is the possibility to implement complex control systems without compromising the local processing of the equipment.

Automatically dividing an image into regions of similar properties, named segmentation, is a challenging task for computers, and it can avoid human errors induced by fatigue. One area that may be greatly benefited from automatic segmentation methods is medical imaging analysis. Within it, chest X-rays are amongst the cheapest and most widely available type of medical images. Provided they can be used for diagnosing lung related diseases, they are an excellent target for automatic image segmentation methods. The current state-of-the art image segmentation relies on manual labels defined a priori to ’learn’ the necessary features for this task. Deep unsupervised learning stands as an interesting alternative to supervised methods, since it only requires the input (e.g. the image X-ray) for training. Due to the visual nature of image segmentation and the standout aspect of the lungs on an X-ray, the combination of unsupervised learning and visual saliency (i.e. the attempt to model human visual attention) is tested for the lung segmentation on X-ray images. The saliency method is compared to state-of-the-art supervised and unsupervised models designed for grayscale medical image segmentation. Results using the Dice, jaccard, precision and recall scores on JSRT and MC datasets indicate that the saliency method enhanced performance over other unsupervised approaches is statistically significant. When compared to supervised models, the saliency method appears to adequately substitute them given the flexibility achieved by the independence from manual labels. Future work includes segmenting the cardiac area and identifying anomalies on X-ray images in an unsupervised fashion.

In this master's study, we address the complexities and implications of cancer, an intricate disease with a significant impact on global health. Our focus is directed towards three critical medical conditions: liver cancer, lung cancer, and cardiac arrhythmia. The core of this project lies in the development of ARFACTA, an innovative biomedical device that utilizes radiofrequency ablation as a treatment. Additionally, we investigate the characterization of roll-off curves in post-mortem bovine and porcine tissue maintained in an ex-vivo state. This research meets the growing demand for affordable treatments within the Unified Health System (SUS), based on data that highlight the costliness of import alternatives. The primary goal is to improve the quality of life for patients, reduce the costs associated with treatment, and alleviate the strain on healthcare systems, both public and private.

Historically, each Industrial Revolution emerged supported by technologies that increased efficiency and reduced costs. Industry 4.0, acclaimed by many as the 4th Industrial Revolution, is supported by a set of Enabling Technologies, such as Big data, Robotics, Artificial Intelligence, Cloud Computing, Additive Manufacturing, Internet of Things, and Virtual/Augmented Reality, among others. When associating these technologies, there is a convergence to the Cyber-Physical Systems concept, and one of its implementations is the Digital Twin (DT). A DT is commonly defined as a virtual representation of a physical object or process, allowing ultra-realistic simulations of physical models, capturing historical data, and real-time processing/monitoring. Different levels of abstraction and detail offer varied views of industrial systems, equipment, processes, and products. Given its complexity and multi-technological composition, its implementation is not smooth. In this context, we propose using a methodology based on Model-Based Systems Engineering (MBSE) to support a model industry's technological evolution, aiming to implement its DT. Our model industry is an innovation laboratory comprising several sectors: Management, Design Office, Manufacturing (several processes), Stock, and Maintenance, among others. In this work, we focus on the area of Additive Manufacturing for external customers (private companies) and internal ones (academic projects), providing the modelling of the optimal path from the arrival of a work order to 3D printing and final delivery of parts. Using the Reference Architecture Model for Industry 4.0 (RAMI-4.0) and the ISO Digital Twin Framework for Manufacturing (ISO-2347), we modelled the system at several levels with the MBSE Arcadia method through the free tool Capella. As a result, we reached the Final System Architecture, indicating the actors involved in the scenario, their functionalities and the interaction flow among all the components. This model allows the simulation and analysis in a single framework considering the complete system and not only isolated subsystems. It provides us with correlations among indicators of different abstraction levels, generating the basis for implementing a multi-scope DT. We can highlight that MBSE systems, even supported by formal methods of Systems Engineering, still cannot be considered mature enough. Consequently, its tools constantly evolve, and new applications favour its consolidation. Therefore, the principles used to construct this work result in an innovative methodology by allying MBSE's tooling support to implement Digital Twins for Manufacturing Systems efficiently.

In recent years, massive Multiple-Input-Multiple-Output (MIMO) systems have been the subject of intense research due to their great potential to provide energy efficiency, data rate gains and diversity through the transmission and reception of multiple versions of the same signal for multiusers. However, its performance benefits depend heavily on the channel estimate at the base station and the way the symbol arrays are transmitted. For modeling and estimating received signals, channel modeling and information transmission (symbol matrices) are very important. In estimating the MIMO channel, it is important to accurately estimate all the parameters that model the channel, such as azimuth and elevation angles, path directions, polarization and amplitude parameters on both sides of the link. For the transmitted symbols it is important to estimate the information received as accurately as possible, for this, in the transmission of the symbols in most cases we use spatial, temporal or frequency coding to increase the diversity of the systems. These codings generally improve the performance, the reliability of the systems and a better estimate of the received signals. In this context, in recent years, semi-blind receivers based on tensor decompositions for MIMO massive systems have been extensively studied. These receivers allow us a better estimate of the channel and symbols without any information about the channel. In some cases, in addition to estimating the channel, it is possible to estimate the channel parameters. This thesis presents received signal model based on tensor decompositions that combine a extension of the MKronST coding and 5th-order channel tensor to transmit the symbols. The coding extension is based on the combination of the tensor space time (TST) coding and the multiple Kronecker product of symbol matrices, called TST-MSMKron coding. The channel assumes a uniform rectangular array (URA) at both, transmitter and receiver which allows us to model the channel as a tensor. More specifically, the theoretical contributions of this thesis are around the proposal of new semi-blind receivers to jointly estimate the symbol matrices, channel and channel parameters without prior knowledge of the channel and channel parameters. What makes this semi-blind estimation possible is the use of the TST-MSMKron coding to transmit the signal. In the first part of this thesis, a multidimensional CX decomposition for tensors is proposed. CX Decomposition is applied to data reconstruction. Based on CX model, one algorithm is proposed to estimate and reconstruct the data tensor. In the second part of this thesis, the TST- MSMKron coding is presented for massive MIMO systems, where a model of the received signal is proposed that combines a 5th-order channel with the TST-MSMKron coding. This system allows us to model the received signal as a coupled-nested-TuckerPARAFAC. In addition, semi-blind receivers in two steps are proposed to jointly estimate the symbols, the channel and the channel parameters. Conditions related to the identifiability and the computational complexity of the proposed algorithms are also discussed in both parts of the thesis. In each part of the thesis, results from Monte Carlo simulations are provided to evaluate the performance of the proposed algorithms. Results show the efficiency of the algorithms in the reconstruction of the datas and joint estimation of the symbols, channel and channel parameters of the system, respectively: In recent years, massive Multiple-Input-Multiple-Output (MIMO) systems have been the subject of intense research due to their great potential to provide energy efficiency, data rate gains and diversity through the transmission and reception of multiple versions of the same signal for multiusers. However, its performance benefits depend heavily on the channel estimate at the base station and the way the symbol arrays are transmitted. For modeling and estimating received signals, channel modeling and information transmission (symbol matrices) are very important. In estimating the MIMO channel, it is important to accurately estimate all the parameters that model the channel, such as azimuth and elevation angles, path directions, polarization and amplitude parameters on both sides of the link. For the transmitted symbols it is important to estimate the information received as accurately as possible, for this, in the transmission of the symbols in most cases we use spatial, temporal or frequency coding to increase the diversity of the systems. These codings generally improve the performance, the reliability of the systems and a better estimate of the received signals. In this context, in recent years, semi-blind receivers based on tensor decompositions for MIMO massive systems have been extensively studied. These receivers allow us a better estimate of the channel and symbols without any information about the channel. In some cases, in addition to estimating the channel, it is possible to estimate the channel parameters. This thesis presents received signal model based on tensor decompositions that combine a extension of the MKronST coding and 5th-order channel tensor to transmit the symbols. The coding extension is based on the combination of the tensor space time (TST) coding and the multiple Kronecker product of symbol matrices, called TST-MSMKron coding. The channel assumes a uniform rectangular array (URA) at both, transmitter and receiver which allows us to model the channel as a tensor. More specifically, the theoretical contributions of this thesis are around the proposal of new semi-blind receivers to jointly estimate the symbol matrices, channel and channel parameters without prior knowledge of the channel and channel parameters. What makes this semi-blind estimation possible is the use of the TST-MSMKron coding to transmit the signal. In the first part of this thesis, a multidimensional CX decomposition for tensors is proposed. CX Decomposition is applied to data reconstruction. Based on CX model, one algorithm is proposed to estimate and reconstruct the data tensor. In the second part of this thesis, the TST- MSMKron coding is presented for massive MIMO systems, where a model of the received signal is proposed that combines a 5th-order channel with the TST-MSMKron coding. This system allows us to model the received signal as a coupled-nested-TuckerPARAFAC. In addition, semi-blind receivers in two steps are proposed to jointly estimate the symbols, the channel and the channel parameters. Conditions related to the identifiability and the computational complexity of the proposed algorithms are also discussed in both parts of the thesis. In each part of the thesis, results from Monte Carlo simulations are provided to evaluate the performance of the proposed algorithms. Results show the efficiency of the algorithms in the reconstruction of the datas and joint estimation of the symbols, channel and channel parameters of the system, respectively

The general objective of this work is to produce ceramic supports based on hydroxyapatite doped with cobalt ferrite and to evaluate the influence of the dopant element on its microstructural characteristics and its porous structure. The ceramic supports were produced by freeze-casting process which created a matrix of interconnected and unidirectional pores. The solids concentration used for the production of the ceramic supports, utilizing freeze-casting was 12.5% vol. Camphene was used as a solvent. Cobalt ferrite was produced using sol-gel and was later used in the doping of hydroxyapatite at 2, 6 and 10% vol. The cooling rate, the dopant content, the sintering temperature and the compressive strength of the ceramic supports were analyzed.. The porosity was measured by the Architect's Principle.Using the X-ray diffraction analysis it was possible to observe the formation of TCP-β for pure hydroxyapatite and under lower conditions of cobalt ferrite. SEM images showed that higher concentrations of cobalt ferrite favored pore growth. Cobalt ferrite contributed to a greater densification of the material, raising the compressive strength from 0.64 MPa to 2.07 MPa in supports cooled to 35 °C and from 0.85 MPa to 2.50 MPa in supports cooled in liquid nitrogen

The ability to learn with few visual samples is one of the main challenges in the field of machine learning. Deep learning has been successful in large data sets applications, but models tend to drastically reduce their performance when dealing with small datasets. Few-shot Learning (FSL), however, is a recent and suitable learning approach to deal with few samples. However, current solutions still seek improvements in the image feature extraction processes to differentiate them effectively. In this work, we conducted a study to provide an improved metric-based FSL classification model. Starting with a literature review, we distinguish the main state-of-the-art models and describe the main approaches with application in insect classification. We explored
the use of Bregman divergences as similarity methods in FSL and investigated the parametric training conditions to minimize classification errors during the test time. The proposed model includes two main modules: 1) a low, middle, and high-level image feature extraction and fusion mechanism (FMC), with an exclusive method of merging these features to generate information-rich representative vectors of classes, and 2) the relative entropy as a promising alternative to deal with the comparison of vectors generated by the FMC in tasks with a few data. The model was tested using two challenging agricultural insect datasets proposed in this work: the first containing insect pests separated by stages of maturity, and the second covering pests and beneficial insects of the maize crops. A wide experimental chain was carried out for the divergence analysis and validation of the proposed model. Validation was performed by comparing our results to those of convolutional neural networks consolidated in the literature. The results showed that the presented model improved the accuracy by up to 3.62% and 2% in 1-shot and 5shot classification tasks, respectively, compared to the traditional global image feature extraction model. Furthermore, the model outperformed the results of the ResNet50, VGG16, and Mobilenetv2 networks, with the advantage of reducing the number of learnable parameters by up to 99%. Few-shot learning is a relevant approach for research related to insect classification in agriculture, serving as an
alternative for developing fast and accurate solutions for application in the crop field.

Machine-learning methods depend heavily on how well the selected feature extractor can represent the raw input data. Nowadays, we have more data and computational capacity to deal with it. With Convolutional Neural Networks, we have a network that is easier to train and generalizes much better than usual. However, a good amount of essential features are discarded in this process, even when using a powerful CNN. Multistream Convolutional Neural Networks can process more than one input using separate streams and are designed using any classical CNN architecture as a base. The use of M-CNNs generates more features and thus, improves the overall outcome. This work explored M-CNNs architectures and how the stream signals behave during the processing, coming up with a novel M-CNN cross-fusion strategy. The new module is first validated with a standard dataset, CIFAR-10, and compared with the corresponding networks (single-stream CNN and late fusion M-CNN). Early results on this scenario showed that our adapted model outperformed all the abovementioned models by at least 28% compared to all tested models. Expanding the test, we used the backbones of former state-of-the-art networks on image classification and additional datasets to investigate if the technique can put these designs back in the game. On the NORB dataset, we showed that we could increase accuracy up to 63.21% compared to basic M-CNNs structures. Varying our applications, the mAP@75 of the BDD100K multi-object detection and recognition dataset improved by 50.16% compared to its unadapted version, even when trained from scratch. The proposed fusion demonstrated robustness and stability, even when distractors were used as inputs. While our goal is to reuse previous state-of-the-art architectures with few modifications, we also expose the disadvantages of our explored strategy.

This thesis presents the application of the Boundary Element Method in the development of simulators for the study of water and gas coning phenomena in horizontal and vertical oil wells. The scope is the two-dimensional and axisymmetric numerical modeling of the phenomena using the Boundary Element Method with sub-regions and interface between fluids modeled as a moving boundary. This text provides a review about these phenomena and their impact on the petroleum industry, in order to situate the reader on the motivation for the choice of the application. Two potential flow models are used, single-phase and two-phase. A third model, single-phase diffusive flow, is also applied. In addition to these, single-phase and two-phase models are adopted for potential flow in axisymmetric domains. All the mathematical treatment involved in the deduction of the governing equations of the models is presented. The mathematical developments of several formulations of the Boundary Element Method are also presented. These are the formulation for two-dimensional potential problems, the Dual Reciprocity formulation for problems governed by the diffusivity equation in a two-dimensional domain, the formulation for potential problems in a three-dimensional axisymmetric domain, and the subregion formulation, for problems over piecewise homogeneous domains. An alternative in modeling the horizontal oil well in two-dimensional domain is proposed, using a set of point sinks to represent its perimeter. This approach allows satisfactory analysis of the behavior of the fluid interface, even after it has already touched the well. The main aspects of the simulators implementation are documented in a way that allows this text to be used as a reference for similar implementations. The determination and application of the boundary conditions for all models, including the compatibility and equilibrium conditions at the fluid interface, are presented. The proposed simulators have been developed and validated against analytical, numerical, and experimental results