| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 16.09 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
Esta dissertação tem como objetivo abordar o desafio das falhas em Sistemas Ciberfísicos (CPSs), que são sistemas integrados e combinam componentes físicos e computacionais. Os CPSs são essenciais em várias áreas, erros nesses sistemas podem causar problemas significativos, como interrupções de serviços, perda de dados ou até mesmo ameaças à segurança. Com base nesse contexto, o objetivo deste estudo foi desenvolver um protótipo para detectar falhas em um CPS laboratorial. Especificamente, o foco foi direcionado para a detecção de falhas em baterias de Polímero de Lítio (Li-Po), o mesmo tipo de bateria utilizado nos robôs móveis do CPS laboratorial. A estrutura implementada para o protótipo seguiu duas abordagens, a primeira é baseada em uma estrutura central no sistema Robot Operating System (ROS) que recebe todas as informações dos robôs móveis em um único nó ROS e a outra é baseada em uma estrutura local, onde cada robô móvel possui um nó ROS dedicado. Para enfrentar esse desafio, foram aplicadas técnicas de Aprendizado de Máquina (ML) do tipo supervisionado e não supervisionado, visando comparar qual tipo desempenharia melhor. Para a estrutura central, os modelos de ML foram treinados utilizando dados generalizados, já para a estrutura local, os modelos de ML foram treinados utilizando dados específicos de cada robô. Essas abordagens permitem que o sistema aprenda a reconhecer padrões e associações nos dados coletados, possibilitando a detecção automática de falhas. Foram coletados dados em relação ao comportamento das baterias de Li-Po por meio de experimentos em laboratório, simulando duas condições de operação, a primeira é relacionada à descarga da bateria quando o robô do CPS se encontra parado e a segunda é quando o robô está em movimento. Esses dados foram preparados para treinar e validar os modelos de ML. Os modelos de ML treinados foram introduzidos nos protótipos de
estrutura central e local, o desempenho deles foi avaliado utilizando dados fornecidos por um ambiente de simulação no CoppeliaSim, onde foi simulado dois tipos de falhas que podem ocorrer nas baterias de Li-Po.
O desempenho dos modelos de ML aplicados e dos protótipos de detecção de falha foi medido através de métricas como acurácia e F1-score. Os resultados obtidos demonstraram que os modelos de ML supervisionados desempenharam melhor do que os não supervisionados para esta abordagem, assim como o protótipo com estrutura local se sobressaiu em relação ao protótipo com estrutura central. Em especial o modelo de aprendizagem de máquina Floresta Aleatória (RF) utilizado no protótipo de estrutura local, sendo capaz de detectar as falhas nas baterias com uma acurácia de 98,8% e F1-score(0) (relacionado a classe dos valores consistentes) de 1, e F1-score(1) relacionado a classe dos valores inconsistentes) de aproximadamente 0,68. Isso indica que o protótipo com estrutura local utilizando o modelo de aprendizagem de máquina RF pode ser efetivamente aplicado para identificar problemas em baterias de Li-Po e, potencialmente, em outros componentes do CPS.
Em conclusão, este estudo contribuiu para a área de detecção de falhas em CPSs. Os métodos utilizados baseados em modelos de ML supervisionados demonstraram ser eficientes na detecção de falhas em baterias de Li-Po em um ambiente laboratorial.
This dissertation aims to address the challenge of failures in Cyber-Physical Systems (CPSs), which are integrated systems that combine physical and computational components. CPSs are essential in various areas, and errors in these systems can cause significant problems such as service interruptions, data loss, or even security threats. Based on this context, the objective of this study was to develop a prototype for detecting failures in a laboratory CPS. Specifically, the focus was on detecting failures in Lithium Polymer (Li-Po) batteries, the same type of battery used in the mobile robots of the laboratory CPS. The implemented structure for the prototype followed two approaches. The first approach is based on a centralized framework in the Robot Operating System (ROS), which receives all the information from the mobile robots in a single ROS node. The second approach is based on a local framework, where each mobile robot has a dedicated ROS node. To address this challenge, supervised and unsupervised Machine Learning (ML) techniques were applied to compare which type would perform better. For the centralized structure, the ML models were trained using generalized data, while for the local structure, the ML models were trained using specific data from each robot. These approaches allow the system to learn to recognize patterns and associations in the collected data, enabling automatic fault detection. Data regarding the behavior of Li-Po batteries were collected through laboratory experiments, simulating two operating conditions. The first condition is related to battery discharge when the CPS robot is stationary, and the second is when the robot is in motion. These data were prepared to train and validate the ML models. The trained ML models were introduced into the prototypes of the centralized and local structures, and their performance was evaluated using data provided by a simulation environment in CoppeliaSim, where two types of failures that can occur in Li-Po batteries were simulated. The performance of the applied ML models and fault detection prototypes was measured using metrics such as accuracy and F1-Score. The results demonstrated that the supervised ML models performed better than the unsupervised models for this approach, and the local structure prototype outperformed the centralized structure prototype. In particular, the Random Forest (RF) machine learning model used in the local structure prototype was able to detect battery failures with an accuracy of 98.8%, an F1-Score(0) (related to the class of consistent values) of 1, and an F1-Score(1) (related to the class of inconsistent values) of approximately 0.68. This indicates that the local structure prototype using the RF machine learning model can effectively be applied to identify problems in Li-Po batteries and potentially in other components of CPSs. In conclusion, this study contributed to the field of CPS fault detection. The methods used, based on supervised ML models, proved to be efficient in detecting failures in Li-Po batteries in a laboratory environment.
This dissertation aims to address the challenge of failures in Cyber-Physical Systems (CPSs), which are integrated systems that combine physical and computational components. CPSs are essential in various areas, and errors in these systems can cause significant problems such as service interruptions, data loss, or even security threats. Based on this context, the objective of this study was to develop a prototype for detecting failures in a laboratory CPS. Specifically, the focus was on detecting failures in Lithium Polymer (Li-Po) batteries, the same type of battery used in the mobile robots of the laboratory CPS. The implemented structure for the prototype followed two approaches. The first approach is based on a centralized framework in the Robot Operating System (ROS), which receives all the information from the mobile robots in a single ROS node. The second approach is based on a local framework, where each mobile robot has a dedicated ROS node. To address this challenge, supervised and unsupervised Machine Learning (ML) techniques were applied to compare which type would perform better. For the centralized structure, the ML models were trained using generalized data, while for the local structure, the ML models were trained using specific data from each robot. These approaches allow the system to learn to recognize patterns and associations in the collected data, enabling automatic fault detection. Data regarding the behavior of Li-Po batteries were collected through laboratory experiments, simulating two operating conditions. The first condition is related to battery discharge when the CPS robot is stationary, and the second is when the robot is in motion. These data were prepared to train and validate the ML models. The trained ML models were introduced into the prototypes of the centralized and local structures, and their performance was evaluated using data provided by a simulation environment in CoppeliaSim, where two types of failures that can occur in Li-Po batteries were simulated. The performance of the applied ML models and fault detection prototypes was measured using metrics such as accuracy and F1-Score. The results demonstrated that the supervised ML models performed better than the unsupervised models for this approach, and the local structure prototype outperformed the centralized structure prototype. In particular, the Random Forest (RF) machine learning model used in the local structure prototype was able to detect battery failures with an accuracy of 98.8%, an F1-Score(0) (related to the class of consistent values) of 1, and an F1-Score(1) (related to the class of inconsistent values) of approximately 0.68. This indicates that the local structure prototype using the RF machine learning model can effectively be applied to identify problems in Li-Po batteries and potentially in other components of CPSs. In conclusion, this study contributed to the field of CPS fault detection. The methods used, based on supervised ML models, proved to be efficient in detecting failures in Li-Po batteries in a laboratory environment.
Descrição
Mestrado de dupla diplomação com a UTFPR - Universidade Tecnológica Federal do Paraná
Palavras-chave
Sistemas ciberfísicos Detecção e diagnóstico de falhas Aprendizado de máquina