Programa de Pós-Graduação em Engenharia Elétrica e de Computação
Permanent URI for this communityhttps://repositorio.ufrn.br/handle/123456789/11949
Browse
7 results
Search Results
Doctoral Thesis Treinamento consciente de quantização vetorial evolutiva: um método adaptativo para compactar modelos de aprendizado de máquina(Universidade Federal do Rio Grande do Norte, 2026-02-06) Flores, Thommas Kevin Sales; Silva, Ivanovitch Medeiros Dantas da; Costa, Daniel Gouveia; https://orcid.org/0000-0003-3988-8476; http://lattes.cnpq.br/0200709054904002; https://orcid.org/0000-0002-0116-6489; http://lattes.cnpq.br/3608440944832201; https://orcid.org/0000-0003-2808-8529; http://lattes.cnpq.br/0630479458408181; Vieira, Tiago Figueiredo; http://lattes.cnpq.br/8601011832053651; Martins, Allan de Medeiros; https://orcid.org/0000-0002-9486-4509; http://lattes.cnpq.br/4402694969508077; Villanueva, Juan Moisés Mauricio; https://orcid.org/0000-0002-8760-9390; http://lattes.cnpq.br/1715030763709959; Oliveira, Luiz Affonso Henderson Guedes de; https://orcid.org/0000-0003-2690-1563; http://lattes.cnpq.br/7987212907837941The consolidation of Large Language Models poses significant challenges to their deployment on edge devices, as the reliance on cloud-based infrastructures compromises data privacy, increases latency, and limits critical applications in scenarios with unstable connectivity. Traditional compression techniques, predominantly applied after training, prove inadequate in this context, as they may degrade model performance or rely on opaque execution pipelines, hindering the analysis and verification of system behavior. In response to these challenges, this thesis proposes Evolving Vector Quantization-Aware Training (EVQAT), whose central principle consists of integrating compression directly into the learning process. In this approach, quantization is no longer treated as a post-training step but rather as an adaptive mechanism during training, allowing the transformations imposed on model parameters to be explicitly controlled and analyzed. mThe methodology is grounded in evolutionary vector quantization theory by treating the weights of the language model as continuous data streams, enabling incremental clustering algorithms such as AutoCloud, Mean-Shift, and Affinity Propagation to dynamically adjust the codebooks in synchrony with the optimization of high-dimensional parameters, thereby enabling model compression with explicit control over quantization error. The integration of codebook evolution into the training process ensures that the partitioning of the vector space follows the learning dynamics, preserving relevant statistical properties and reducing the impact of quantization noise. This characteristic makes EVQAT suitable for quantizationsensitive architectures such as Transformers, extending its applicability from multilayer neural networks to generative language models under strict memory and computational constraints. The proposed approach is evaluated across multiple experimental scenarios, including training without quantization, quantization-aware training using evolutionary vector quantization and 8-bit integer (INT8) representations, as well as post-training quantization under the same variants, with deployments on microcontroller-based platforms. In this context, auditability is achieved through transparent access to model parameters, including the vocabulary, the weights of each hidden layer, and the attention mechanisms. As a result, the methodology enables the automatic generation of standard, self-contained C++ code, with real-time inference capabilities and direct execution on diverse hardware architectures, without reliance on third-party libraries or operating systems, ensuring explainability at the levels of training, model, and execution, as well as interoperability across platforms.Master Thesis MMCloud: um algoritmo de clustering não supervisionado online para análise do comportamento do motorista(Universidade Federal do Rio Grande do Norte, 2025-07-10) Medeiros, Morsinaldo de Azevedo; Silva, Ivanovitch Medeiros Dantas da; Silva, Marianne Batista Diniz da; https://orcid.org/0000-0002-8277-7571; http://lattes.cnpq.br/6470261020797104; https://orcid.org/0000-0002-0116-6489; http://lattes.cnpq.br/3608440944832201; https://orcid.org/0000-0001-7624-5301; http://lattes.cnpq.br/2910614991697297; Oliveira, Luiz Affonso Henderson Guedes de; https://orcid.org/0000-0003-2690-1563; http://lattes.cnpq.br/7987212907837941; Santos, Max Mauro Dias; http://lattes.cnpq.br/6212006974231025Evolving and adaptive intelligent systems are essential for handling dynamic data streams from real-world environments, especially in resource-constrained settings such as vehicles enabled with the Internet of Things (IoT). This study proposes an online evolutionary clustering approach for real-time driver behavior diagnosis, leveraging Tiny Machine Learning (TinyML) and soft sensors. It introduces the MMCloud algorithm, an incremental clustering model designed to manage continuous data streams. The algorithm eliminates the need for retraining and is capable of adapting to concept drift and varying driving conditions. The methodology integrates onboard diagnostics (OBD-II) with an edge computing framework to classify driver behavior into three categories: cautious, normal, and aggressive. To validate the proposed approach, two case studies were conducted in urban environments under various traffic conditions. One used a Freematics One+ device, while the other deployed the embedded algorithm in a mobile application. The results demonstrate the system’s ability to accurately identify evolving driving patterns, contributing to safer driving practices, fuel consumption optimization, and intelligent transportation systems (ITS). This research advances the field of evolutionary AI by integrating adaptive machine learning models with embedded IoT solutions, enabling autonomous and self-organizing monitoring of driver behavior.Master Thesis Detecção de falhas em operação de bombas elétricas utilizando TinyML e sinais acústicos(Universidade Federal do Rio Grande do Norte, 2024-11-22) Silva, Deivison Luan Xavier; Oliveira, Luiz Affonso Henderson Guedes de; Gendriz, Ignacio Sanchez; https://orcid.org/0000-0003-2690-1563; http://lattes.cnpq.br/7987212907837941; https://orcid.org/0009-0000-2388-0571; http://lattes.cnpq.br/3317633556320520; Bustos, Harold Ivan Ângulo; Silva, Ivanovitch Medeiros Dantas daMost production processes begin with the generation of kinetic energy by electric motors. In fluid-related operations, electric pumps are responsible for the movement and pressure variation of fluids. Given the importance of this equipment, failures in their operation can negatively impact the entire production chain, causing reduced efficiency, unforeseen stoppages, and potential safety risks for workers. To mitigate these issues, industrial asset monitoring strategies are implemented, assessing factors such as the condition and performance of pumps, aiding in corrective and preventive maintenance decisions. Traditionally, vibration and current sensors are used to monitor the condition of this equipment, but due to cost and the need for physical contact, their application may be unfeasible in some cases. Acoustic sensors are a low-impact alternative, capable of detecting anomalies at early stages without physical contact, offering an efficient and non-intrusive monitoring solution. For detecting and diagnosing defects in electric pumps, machine learning techniques, especially neural networks, are widely used. This study investigates the feasibility of using miniaturized machine learning, known as TinyML, in combination with acoustic sensors for detecting defects in electric pumps. Specifically, an audio classification model based on convolutional neural networks (CNN) was developed, implemented, and embedded on the Arduino Nano platform. To validate the proposal, the "MIMII Dataset: Sound Dataset for Malfunctioning Industrial Machine Investigation and Inspection" was used, containing audio recordings of defects from different pump models. To ensure the applicability of the embedded system, the relationship between neural network model complexity and its accuracy in detecting anomalies was analyzed, using quantization techniques to reduce the CNN by approximately 50%. Enabling the creation, training, and deployment of machine learning models on edge devices, the Edge Impulse platform was used for developing the fault detection model using acoustic sensors and its compression into a reduced model optimized for the Arduino. The results indicate an accuracy of over 96% in defect detection, providing a low-cost, highly efficient, and easy-to-implement solution with potential applications in various industrial areas, significantly improving asset monitoring and reducing maintenance-related costs.Master Thesis An evolving multivariate time series compression algorithm for IoT applications(Universidade Federal do Rio Grande do Norte, 2024-12-19) Costa, Hagi Jakobson Dantas da; Silva, Ivanovitch Medeiros Dantas da; Silva, Marianne Batista Diniz da; https://orcid.org/0000-0002-0116-6489; http://lattes.cnpq.br/3608440944832201; http://lattes.cnpq.br/6563621187728660; Viegas, Carlos Manuel Dias; Gendriz, Ignacio SanchezThe Internet of Things (IoT) is transforming how devices interact and share data, especially in areas like vehicle monitoring. However, transmitting large volumes of real-time data can result in high latency and substantial energy consumption. In this context, Tiny Machine Learning (TinyML) emerges as a promising solution, enabling the execution of machine-learning models on resource-constrained embedded devices. This paper aims to develop two online multivariate time series compression approaches specifically designed for TinyML, utilizing the Typicality and Eccentricity Data Analytics (TEDA) framework. The proposed approaches are based on data eccentricity and do not require predefined mathematical models or assumptions about data distribution, thereby optimizing compression performance. Both approaches were applied to two case studies: one using the Freematics ONE+ device for vehicle monitoring in an embedded scenario, and another using the OBD-II dataset collected from the Freematics ONE+ in a non-embedded context. The results indicate that the proposed approaches, whether for parallel or sequential compression, present significant improvements in runtime and compression errors. These findings highlight the potential of the approaches to improve the performance of embedded IoT systems, enhancing the efficiency and sustainability of vehicular applications.Doctoral Thesis An unsupervised tinyML incremental learning approach for outlier processing and forecasting(Universidade Federal do Rio Grande do Norte, 2024-08-30) Andrade, Pedro Henrique Meira de; https://orcid.org/0000-0002-0116-6489; http://lattes.cnpq.br/3608440944832201; https://orcid.org/0000-0002-7729-9085; http://lattes.cnpq.br/6695123583643731; Oliveira, Luiz Affonso Henderson Guedes de; Barros, Tiago Tavares Leite; Costa, Daniel Gouveia; Villanueva, Juan Moisés MauricioThe Internet of Things (IoT) is a paradigm where computing and connectivity capabilities are embedded into objects, connecting them to the Internet. Acknowledged as a crucial and emerging technological area, IoT holds significant potential to enhance the quality of life, optimize industrial processes, and offer more applications to everyday objects. With the increasing number of IoT-connected devices, there arises a need for infrastructure to manage the vast volume of generated data. In this context, Edge Computing stands out by processing data close to its source, leaving only heavier processing tasks for central servers. Edge processing enables the development of optimized machine learning algorithms, known as Tiny Machine Learning (TinyML). By employing lightweight and optimized algorithms, TinyML offers advantages such as reduced latency, improved energy efficiency, and increased autonomy for devices operating in remote or isolated applications. In the field of TinyML, implementing machine learning techniques on resource-constrained devices like microcontrollers poses significant challenges, including outlier detection and correction. This work contributes to developing an unsupervised incremental learning algorithm for outlier processing within the context of TinyML. This innovative approach applies unsupervised machine learning for outlier detection and correction on resource-constrained devices, adapting to external variations over time. The algorithm addresses the problem of signal processing at the edge of IoT applications, enabling, for example, a smart meter to process events locally before sending data to the supervisory system. The solution was implemented and validated through simulations and tested on two different microcontrollers: the ATmega328P (Arduino) and the Espressif ESP32 (Freematics), confirming its feasibility and good performance. This work fills a gap in the literature by introducing a new approach for data processing on resource-limited devices, utilizing an incremental learning technique. The evaluation compared the results obtained on embedded systems with those obtained on computers using different programming languages and tools.Master Thesis Uma proposta para detecção de buracos com aprendizagem de máquina na borda(Universidade Federal do Rio Grande do Norte, 2024-05-14) Silva, Jordão Paulino Cassiano da; Silva, Ivanovitch Medeiros Dantas da; https://orcid.org/0000-0002-0116-6489; http://lattes.cnpq.br/3608440944832201; https://orcid.org/0000-0003-1691-3288; http://lattes.cnpq.br/1066846658119458; Oliveira, Luiz Affonso Henderson Guedes de; https://orcid.org/0000-0003-2690-1563; http://lattes.cnpq.br/7987212907837941; Costa, Daniel GouveiaPotholes in urban roads represent a significant problem, affecting both user safety and vehicle durability. This study addresses the urgent need for effective solutions for pothole detection that can be implemented in real-time using devices with limited computational resources. An innovative approach was developed, integrating machine learning on edge devices, with an emphasis on YOLOv8 and FOMO models within the TinyML context. A specialized dataset, containing annotated images, was used to train these models for accurate pothole detection. The optimization of YOLOv8 and FOMO models’ performance for edge devices ensures real-time efficiency. This work not only provides effectively trained models but also presents an adaptable framework for pothole detection, ensuring practical and efficient implementation. Additionally, a complete pipeline for pothole detection is proposed, validating the models’ accuracy and efficiency. This approach offers a robust solution for the automatic recognition of potholes, significantly contributing to improvements in urban infrastructure maintenance.Master Thesis Um algoritmo online e evolutivo para compressão automática de dados em cenários de IoT(Universidade Federal do Rio Grande do Norte, 2021-07-23) Signoretti, Gabriel Lucas Albuquerque Maia; Silva, Ivanovitch Medeiros Dantas da; http://lattes.cnpq.br/3608440944832201; http://lattes.cnpq.br/3335383471638892; Leitão, Gustavo Bezerra Paz; http://lattes.cnpq.br/6766556258983738; Villanueva, Juan Moises Mauricio; http://lattes.cnpq.br/1446817462218646; Oliveira, Luiz Affonso Henderson Guedes de; http://lattes.cnpq.br/7987212907837941With the advancement and mass adoption of solutions in the fields of Internet of Things (IoT) and connected cities, the number of devices and sensors connected to the network tends to grow exponentially. In this scenario, the transmission and storage of the growing volume of data bring new challenges. When devices transmit potentially irrelevant or redundant data, there is increased energy and processing waste, as well as unnecessary use of the communication channel. Thus, local data compression solutions on the IoT devices themselves become increasingly attractive, enabling the elimination of samples that would have little or no contribution to the application, in order to significantly reduce the volume of data needed to represent the information. However, such devices present on the market today have serious storage and processing power limitations. In order to circumvent these limitations, the TinyML field emerges, which seeks ways to implement machine learning models in low-power devices. Given this context, one of the sectors that can benefit most from these new technologies is the automobile industry, as currently all cars produced must be instrumented with a series of sensors. In this way, by connecting an intelligent device to the vehicle, it is possible to process the data locally and transmit it to a remote server later. In this context, the present work proposes the development of a new online, unsupervised, and automatically adaptable data compression algorithm for IoT applications. The proposed approach is called Tiny Anomaly Compressor (TAC) and is based on data eccentricity and does not require pre-established mathematical models or any assumptions about data distribution. To test the effectiveness of the solution and validate it, two tests were carried out with different objectives. First, a comparative analysis on two real-world datasets was developed with two other algorithms from the literature, the Swing Door Trending (SDT) and the Discrete Cosine Transform (DCT). Finally, the proposal was embedded in an IoT device based on an Arduino and connected to a car to verify the impact of the algorithm on the processing time of the system’s primary operations. The obtained results show that it is possible to achieve high compression rates without significant impacts on the generated error and system processing times.